Files
PaperClipAI/doc/run-log-events.md
T
DottaandClaude Opus 5.5 d4498ae7ac fix(runner): settle every recorded tool result and free sessions blocked by a closed turn's pending call (#21490)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - Agents call Paperclip tools through the Runner. The Runner records
each call and returns the controller's result to the provider (here
OpenCode)
> - In chat-eval cohort 13, a `list_workflows` result was recorded but
never reached OpenCode. OpenCode waited about 290 s, and the call stayed
pending
> - The next run's `run.attach` then failed with
`durable_pending_tool_calls`, so the person's follow-up question never
got an answer
> - This pull request fixes the result shape that caused the rejection,
makes the Runner settle every recorded result, and lets the next run
settle a call that a closed turn left pending
> - The benefit is that a tool result always reaches the agent, and a
stuck call can no longer block every later turn

## Linked Issues or Issue Description

**What happened?**
Chat-eval cohort 13 (OpenCode,
`openrouter/deepseek/deepseek-v4.1-flash-1017`, master `df1b12ad8e`). In
`cases.case-history` and `cases.create-case`, the CEO called
`list_documents` (or `list_projects`) and `list_workflows` (call id 5)
together. The server answered `GET /companies/{id}/workflows` with 200
in 584 ms. The OpenCode MCP bridge diagnostic for call 5 says `outcome:
"unanswered"`, `written: false`, `handlerMs: 289038` (290723 in
create-case). The sibling call was answered in about 1 s. The run ended
with `runner_settlement_incomplete`: `semanticTools.pending` held call 5
and its `semantic_tool.result` delivery command had status `failed`. The
next question started a run whose `run.attach` was refused: `run.attach
requires a settled Codex provider session: durable_pending_tool_calls`.
The run failed in 17 s, and the question was not answered.

Cause: `executeNativeWorkflowTool` returned the Runner API result
unchanged: `{ ok: true, status: 200, operationId: "GET
/api/companies/{companyId}/workflows", ... }`. Every other API-backed
tool renames that field to `apiOperationId`. The Rust tool bridge
(`semantic_response_value`) reads an object with a boolean `ok` and an
`operationId` as a semantic envelope that must name this tool and call.
The REST route did not match, so `apply_result` failed with `semantic
result envelope does not match its provider call`. Nothing was sent to
the provider, and the call stayed pending forever. The red test
reproduces this exact error. The parallel call is not the trigger: the
sibling call settled normally, and the OpenCode MCP bridge answered
other parallel pairs in the same runs. `list_workflows` first appears in
cohort 13, so cohorts 11 and 12 did not hit it. All four workflow tools
had this shape.

**Expected behavior**
Every recorded tool result reaches the provider. When a recorded result
cannot be delivered, the provider gets an error at once and the call
settles. A call that a closed turn left pending does not block the next
run.

**Steps to reproduce**
1. Run an OpenCode agent with Cases enabled on the durable Runner.
2. Ask a question that makes the agent call `list_workflows`.
3. See the call wait about 290 s. Then ask a follow-up question and see
`run.attach` fail with `durable_pending_tool_calls`.

**Paperclip version or commit**
`master` at `df1b12ad8e`.

## What Changed

- `server/src/services/native-runtime/native-workflow-tools.ts`:
workflow tool results name the route `apiOperationId`, like case tools
and `call_api`.
- `provider_bridge.rs`: add `settle_undeliverable_result`. When the
bridge rejects a recorded result for a call it still holds, the call
settles with a bounded, non-retryable runner error. Conflicting or
unknown results still fail as before.
- `provider_backend.rs` (`deliver_semantic_result`): use that
settlement, deliver the error to the provider, and record a
`semantic_tool_result_undeliverable` diagnostic with the call,
operation, reason and the rejected result's digest.
- `provider_bridge.rs` and `provider_backend.rs` (`attach_run`): add
`expire_closed_turn_calls`. When nothing is live (no active or ambiguous
turn, no provider request waiting), `run.attach` settles calls that a
closed turn left pending as an expired failure with an unknown outcome,
and records `semantic_tool_result_expired`. The provider stopped waiting
when its turn ended, and the new run's controller never delivers the old
run's results. A call that a live turn still waits on still blocks
attachment. This follows the existing `tool_call_interrupted_by_restart`
rule: never execute the operation again, and record an unknown outcome
when no result can ever arrive.
- Docs: `doc/run-log-events.md` and
`packages/paperclip-runner/docs/durable-recovery.md` describe both
diagnostics.

## Verification

- Red/green, Rust (`cd packages/paperclip-runner/runner && cargo test -p
paperclip-runner-core --test codex_provider --
a_rejected_recorded_result run_attach_expires`):
- Before:
`a_rejected_recorded_result_still_answers_the_provider_and_frees_the_session`
fails with `semantic result envelope does not match its provider call`
(the production error).
`run_attach_expires_a_call_left_pending_by_a_closed_turn` fails with
`run.attach requires a settled Codex provider session:
durable_pending_tool_calls` (the production refusal).
- After: both pass. The fake provider logs `tool-response:failure`, so
it got an answer at once. The next `run.attach` succeeds.
- Parallel calls (`cargo test -p paperclip-runner-core --lib --
provider_bridge`):
`a_rejected_result_in_a_parallel_pair_still_settles_its_own_call` sends
the cohort 13 pair (`list_documents` + `list_workflows`). The rejected
call settles, its sibling keeps its own result, and the next turn
starts. `only_a_closed_turn_expires_its_pending_calls` checks that an
open turn expires nothing.
- Upgrade and mid-run cutover:
`run_attach_expires_a_call_left_pending_by_a_closed_turn` saves state
with a pending call after the turn ends, stops that Runner, and attaches
with a new Runner process from the saved state.
`durable_backend_settles_pending_tools_when_recovery_finds_the_turn_ended`
(unchanged) still passes: a late result in the same run still settles as
`settled_after_turn`.
- Red/green, server (`cd server && npx vitest run
src/services/native-runtime/native-case-tools.integration.test.ts -t
workflows`): before, `expected { ok: true, status: 201, … } to not have
property "operationId"`; after, 2 passed.
- `cargo test -p paperclip-runner-core`: 643 lib tests pass. One lib
test (`fenced_receipt_output_keeps_prefix_when_stdout_never_closes`)
failed once with `EPERM` while signaling a process group under parallel
load. It passes alone 3 of 3 and does not touch this change. The
integration test binaries pass.
- Paid narrow run for `cases.case-history` and `cases.create-case`: see
the comment below.

## Risks

- Previous-release story: no persisted shape changes. State saved by the
previous Runner with a call pending after its turn closed carries
forward. The new Runner settles that call at the next `run.attach`, and
the session continues (tested from saved state with a new process). A
previous-release Runner that receives the corrected `apiOperationId`
result delivers it normally, because the result is no longer read as an
envelope.
- Behavior change: a call left by a closed turn is now settled at the
next run's attachment, not kept forever. Its outcome is recorded as
unknown, and the operation is never run again. A provider turn that is
still live keeps the old refusal (`durable_pending_tool_calls`,
`provider_side_effect_pending`), and the controller still never replaces
the session.
- A recorded result that the bridge rejects now becomes an error for the
agent, not an endless wait. The agent is told the action may have
completed and must check the current state before repeating it.
- Latency: nothing is added before turn start. The attach-time check
walks the in-memory pending calls (usually none) and adds no I/O beyond
the existing state save. The rejected-delivery path now answers in the
same command instead of after the provider's ~290 s client timeout, and
the follow-up run no longer fails after 17 s.

## Model Used

- Claude (Anthropic), model ID `claude-opus-5-5`, used through Claude
Code with tool use and code execution.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

Refs #21398 (first release where `list_workflows` was called in this
eval), #19733 (the attach refusal), #20886 (earlier OpenCode delivery
fix).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-14 14:24:21 -06:00

1264 lines
77 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Run-Log Events
Run-log events write to the `heartbeat_run_events` table
(`packages/db/src/schema/heartbeat_run_events.ts:6-20`). They are not
Paperclip Telemetry events, and they are not OpenTelemetry exports. A run-log
event needs no operator endpoint.
## ACPX sidecar shutdown
The ACPX sidecar emits `runtime.diagnostic` with code
`sidecar_shutdown_requested` once at the first local shutdown request, before
waiters are removed. Its JSON message records the local reason, pending tool,
input and permission counts, completion-normalization count, and at most 16
oldest pending tools. Each tool sample contains elapsed milliseconds and a
SHA-256 call ID. Hash the corresponding `runtime.tool_called` call ID to match
it. Arguments, results, operation names and raw call IDs are excluded.
The existing diagnostic path also writes this message to sidecar stderr.
A controller that has already stopped may not ingest the frame. Stdin EOF
identifies a closed controller pipe, not the upstream reason for that closure.
This is local run evidence; no first-party Telemetry event or OpenTelemetry
span is added. Retry, cancellation and authorization behavior stay unchanged.
## Startup reuse evidence
`run.startup.reuse` uses schema `paperclip.startup-reuse.v1`. Each receipt names
one layer: `environment` (the lease), `workspace`, `runner`, `provider_process`,
or `provider_session`. Its state is `reused`, `fresh`, `unknown`, or
`not_applicable`. Reasons use a closed set: `verified_reuse`, `created`,
`replacement`, `not_found`, `expired`, `identity_mismatch`, `resume_failed`,
`configuration_changed`, `tools_changed`, `workspace_changed`, `cache_miss`,
`not_applicable`, and `evidence_missing`. No paths, identifiers, configuration
values, or exception messages enter these fields.
A restart's reattachment to a remote Runner resumes the run's own retained
sandbox lease and never acquires another one. Its `environment` receipt is
`reused` / `verified_reuse`. Earlier releases reported the lease's original
acquisition there, so a reattachment to a sandbox the run created showed
`fresh` / `created` although no sandbox was created.
Provider-process receipts carry a cumulative `providerLaunchCount` for the run.
`null` means unobserved; zero requires a physical launch observer. Direct ACPX
processes count initial launches and respawns, including launches during retries.
On the remote process-session bridge, ACPX observes the host relay rather than
the sandbox provider. Provider-process reuse and launch counts stay unknown on
that transport. Native Runner session events establish session continuity, but
do not establish provider-process reuse or its launch count. Those values stay
unknown without physical evidence. The final ACPX receipt does not wait for
persistence, so a stalled diagnostic sink cannot hold the staging lease.
The existing successful `runner.process.launch` performance receipt establishes
a fresh Runner process, independently of provider-session continuity.
The Product E2E result stores `startupReuse` by run, and its dashboard displays
each layer, its reason, and the launch count. `warm` requires all applicable
layers to be reused and zero observed provider launches. `cold` requires all
applicable layers to be fresh. Some reuse with other fresh or missing evidence
is `partial_reuse`; otherwise the label is `unknown`. Fresh decisions dominate
later reuse receipts within a run. Legacy runs without receipts retain unknown
layers and a null launch count. Session IDs and follow-up turns never establish
warmth. These events stay in the local run log; they add no first-party Telemetry.
## Ordinary event persistence
Events without a native source identity or retry-exhaustion receipt use one
atomic PostgreSQL statement. The statement checks the stored run's company and
agent, allocates its next sequence, and inserts the event. An insertion failure
also rolls back the sequence allocation. The returned row uses the existing
column decoders, including the lossless payload codec. Callers still wait for
the durable write before they publish or acknowledge the event. Native-source
replay and retry-exhaustion deduplication retain their existing transaction path.
## Legacy case handoff cleanup
`legacy.case_handoff_cleanup_intent` records local cleanup ownership before
dispatch, then binds it atomically to the observed process identity on spawn.
It includes the host, controller/run generation, PID/group/start identity, and
owned scratch descriptor or explicit `null`. It does not prove termination.
`legacy.case_handoff_cleanup_verified` records the later stop proof before
scratch removal. Restart recovery rechecks process, controller, workspace and
ownership gates before completing cleanup and releasing the handoff.
These controller receipts are cleanup bookkeeping, not evidence of agent task
progress, and are excluded from run-liveness action counts.
The controller authors both events on the system stream. Adapter ingestion
rejects the reserved `legacy.case_handoff_` namespace. These events stay in the
instance run log and add no first-party Telemetry or OpenTelemetry export.
## Provider-pack preparation timing
Every `run.performance.span` payload also carries `wallStartedAtMs`, the span's own
start time in epoch milliseconds (since 2026-11-09). Span rows can be written late
through deferred diagnostics, so readers take a span's start from this field (or
from the trace start plus `startOffsetMs`), never from the row's write time.
The preparation spans of a native run (`heartbeat.queue`, `heartbeat.prepare_*`,
`environment.*`) are always written this way: in the order they ended, while the
provider starts, and before the run's trace finishes. They can appear after the
first Runner events of the run.
`run.performance.span` also records `task.time_to_first_token` once per native
dispatch at the first newly committed deliverable text: assistant or reasoning
`item.delta`/`item.completed`, a redaction tail released at completion, or an
accepted run-result summary. Its
`durationMs` starts at the earliest accepted input. The accepted comment or
answered question is the start only when its span is the earliest preparation span.
A comment coalesced into an already-queued run starts after run creation, so that
run uses queued run creation. The durable wake request's `requested_at` replaces
the start when it is earlier (deferred, coalesced and retried wakes). Without
preparation history, the start is dispatch and the anchor is `run_created`. It ends at Core commit callback. `anchor` names the start:
`comment`, `question_response`, `wake_request` or `run_created`. `startOffsetMs` is
zero, or negative when a wake request precedes the run trace. `eventKind` is
`agentMessage` or `reasoning`. The wake request time is read after the first text,
in detached diagnostic work bounded to 2 s (a
transaction-local `statement_timeout`, so PostgreSQL cancels a stalled read), so the measurement adds no wait before
`turn.start` and never delays event delivery. The span can appear after later
run-log rows; the trace settles it before `task.provider_session.measured`. It includes preparation and startup, but excludes browser
delivery/rendering. This measures text availability, not an exact model token.
Empty text, item starts, tools, usage, result proposals, and duplicate replay do
not qualify. No deliverable text means no timing sample. A recovered dispatch can
record newly committed output from the original start; it does not reproduce
earlier output timing. The event contains no text or provider identifiers.
## Speed budget breaches
`run.performance.budget_exceeded` (schema `paperclip.run-performance-budget.v1`,
level `warn`) records one native run that exceeded its p95 speed budget. The
budgets come from the Phase 0 local baseline and are listed in
`doc/plans/2026-11-07-durable-ownership-and-reconciliation.md` ("Speed budgets").
`metric` is `time_to_first_activity` (`task.time_to_first_agent_activity`:
the first text delta, tool-call start or reasoning item; named `ttft` and
judged on `task.time_to_first_token` before 2026-11-08), `agent_boot` (end of
`heartbeat.queue` to the end of `runner.turn.submit` or `runner.session.resume`)
or `restart_recovery_control_plane` (startup recovery `ready` to the start of the
provider's saved-session load, `provider.session.load`, or to the recovered run's
first provider event when nothing reloads). That is Paperclip's own share of a
recovery: claim, prepare, runtime stage and runner launch. The full recovery
(`ready` to the first provider event) and the provider's session load grow with
session size and with how many sessions resume at once, so they are reported in
the payload as `restartTotalMs` and `providerLoadMs` and are not judged; events
written before 2026-11-09 judged the full latency as `restart_recovery`.
The payload names `adapter` (`codex`, `claude` or `opencode`),
`environment` (`local`), `startKind` (`cold`, `warm` or `unknown`; an unknown
start uses the looser budget), `measuredMs`, `budgetP50Ms`, `budgetP95Ms`,
`percentile` (`p95`), `baseline`, and, for restart recovery, `disposition`.
For `time_to_first_activity`, `measuredMs` counts from run creation: when the input waited for its
run to exist (for example a comment posted while the previous turn was still
running), that wait is excluded from the judgment and reported as `inputWaitMs`.
Each metric is checked at most once per dispatch. A restart-recovery dispatch is
judged only on `restart_recovery_control_plane`, because its first activity and boot span the outage. Time to first text (`task.time_to_first_token`) is reported without a budget. Remote runs and other adapters have
no budget yet. The event is written in detached diagnostic work after the
measurement, so it adds no wait before `turn.start`. It is a run-log row only
and never goes to Telemetry.
The existing `run.performance.span` event also records `provider_pack.select`,
`provider_pack.discover`, `provider_pack.verify`, `provider_pack.verify_preinstalled`,
`provider_pack.link_preinstalled`, `provider_pack.clear_stale`, `provider_pack.upload`,
and `provider_pack.verify_uploaded`. Selection records `stateSource` as `staged`,
`preinstalled`, or `uploaded`. Its closed reason is `verified_pack_reused` or
`cached_packs_unverified`. A verification failure describes an unusable candidate;
it does not identify a particular compatibility mismatch or a failed run.
Harness startup also records `harness_state.inspect`, `harness_state.path_check`,
and `harness_state.launch_material` for remote state inspection and fresh launch
material. These events retain the existing schema and carry no paths or exception strings.
They also become native Observability spans only with an operator OTLP endpoint.
## Local runner startup sub-steps
`run.performance.span` splits a local Runner startup (`runner.session.bootstrap` or
`runner.session.resume`) into three sequential children of `runner.session.startup`:
- `runner.runtime.stage`: controller-side launch preparation before the Runner
process exists (runtime context, provider home, launch profiles; on resume, the
durable-state checks and authority rotation as well).
- `runner.process.launch`: from spawning `paperclip-runnerd` until its first
authenticated connection to the PRP control plane. An exact adoption of a live
Runner records no launch, so this span keeps meaning a fresh Runner process
(see startup reuse above).
- `provider.session.load`: from that authentication until the provider session
identity is confirmed. It covers `run.prepare` and `session.open`, or on resume
the `run.attach` that restores the provider and the recovery probe.
Each carries `mode` (`bootstrap` or `resume`), `target: "local"`, and `outcome`
(`failed` when startup fails inside the step). The Runner transport reports each
milestone synchronously; the controller writes the row as deferred diagnostic
work, so the run-log write never delays `turn.start`. Remote launchers keep their
own `runner.runtime.stage` and `runner.process.launch` spans. These rows stay in
the instance run log; with an operator OTLP endpoint they are also Observability
spans. They are not Telemetry.
## Native PRP Run-Log Events
Provider startup command failures and failed provider turns can carry a
`providerFailure` cause with schema `paperclip.provider-failure.v1`. The cause
retains a normalized category, a bounded and redacted provider explanation, the
startup/turn phase, and available status, code, request ID, or retry time. Provider
adapters extract these fields; arbitrary exception objects, request bodies,
headers, environment variables, and stacks are excluded. Unknown categories retain
their explanation. The cause stays in company-scoped run records and the run log;
it is not added to first-party Telemetry, Observability attributes, or Sentry
failure reports. The reportable run error keeps its existing generic text, except
that an ACPX startup rejection keeps Runner's closed rejection text and its
classification. Startup receipts retain the cause across controller restart before
any provider turn exists.
When a failed ACPX control closes the session before polling, the resulting
`turn.failed` event retains that control's closed-code error. A later closed-session
check does not replace it, and a closed code does not infer a provider explanation.
The terminal still reports `recoverable: false` and no observed provider terminal;
it does not prove termination or authorize replay.
Every ACPX session ensure rejection carries a startup cause. When the provider
returned no ACP error, the cause keeps the rejection's own redacted message with
category `unknown`. A provider can type the failure with ACP's
authentication-required code, or with a known error type or `errorKind` in the ACP
error data. The cause then uses that category, and the rejection classification is
`provider_<category>`, such as `provider_quota_exhausted`, instead of
`session_ensure_failed`.
Completed OpenCode builtin `bash` parts preserve the numeric process exit from
`state.metadata.exit` in `tool.execution.completed.payload.exitCode`. The tool
status stays `completed` even when the command exits with a nonzero code. This
status describes completion of the tool call, not success of the command or the
native run. Metadata adds an exit only when it is an integer. Conflicting supplied
legacy exits stay unknown. Existing numeric `state.exit` values remain supported;
metadata on other tools cannot add process-exit proof.
An edit tool call can carry `fileChanges` on its `tool.execution.*` payload.
Each entry names one workspace-relative `path`, its `operation` (`create`,
`modify`, `delete` or `rename`), `previousPath`, `additions`, `deletions`, a
unified `diff` and `diffTruncated`. The runner drops paths outside the working
directory, redacts secrets in each patch, and bounds a patch to 32 KiB and one
call to 32 files. Counts cover the whole patch even when the patch is shortened.
Codex `fileChange` items produce these rows. Claude Edit and Write calls produce
them from the ACP `diff` content. Codex `turn/diff/updated` also records the
whole-turn diff as `workspace.change.updated` (`paperclip.workspace.diff.v1`).
The task chat shows each patch under its edit row. It ends the turn with the
change set, or with one built from the edit rows when the harness reports no
whole-turn diff.
`auth.credentials.updated` is a runner-produced PRP event for a change to a
supported credential store, including account identity changes and later token
refreshes on the new account. The bundled transport awaits a host commit callback
before acknowledging delivery, including during warm idle and provider cleanup.
The callback writes directly to the bound run through the shared run-log writer.
Its separate credential source namespace avoids collisions with driver-normalized
source sequences and deduplicates retries using the original PRP event identity.
The original envelope is retained in `prpEvent`. Its payload contains only a
provider, an opaque binding ID, a revision, and a detection method. It contains
no credentials, account identity, file path, or credential fingerprint. The
consumer can use the event to read the updated credentials through its existing
private credential channel. This run-log event stays in the instance database.
`native.recovery.execution_reclaimed` records an atomic same-run recovery claim
that restores a released task execution binding. The claim locks the task,
coordinator, and original run in that order. It preserves the assignee and checks
current execution/checkout owners, other live work, dependencies, pause holds,
and stop/successor records before restoring the binding. Attachment authorization
keeps its existing requirement for that exact binding. Remote retry claims also
require the exact saved controller authority and reject a newer task run. The
event does not prove that a remote provider survived or permit provider replay.
Skipped local retry candidates wait 30 seconds before recheck. Remote admission
holds use the existing bounded system-recovery backoff independently of provider
attempts. Permanent ineligibility or policy exhaustion records the terminal
automatic recovery disposition and board action without releasing the retained
execution binding, provider, or checkpoint. These settled holds schedule no
further retry. Repeated blocked visits record only a changed hold reason.
The same durable hold prevents background stale-lock cleanup and queue release
from clearing either task binding until the remote termination receipt contract
is satisfied. A terminal heartbeat status is not a provider-stop receipt.
Undated retries sort first. A failed candidate cannot discard earlier committed
claims. Queued waiters do not block the original execution owner from recovery;
they still wait for that owner to settle before normal queue promotion.
This is local run-log data.
Fresh remote Codex model substitution emits `runner.model_fallback` on the
system stream with `warn` severity. Its payload contains `requestedModel`,
`effectiveModel`, and `codexCliVersion`. It records a preparation choice before
provider launch, not a failed turn or a retry. The same substitution appears as
a system warning in the task conversation. It remains local run-log data.
The hidden native coordinator writes each validated PRP event to the bound
run's existing event stream before it acknowledges the runner. The row keeps
the PRP `eventType`, source instance, source event ID, source sequence, protocol
schema version, and a SHA-256 digest of the canonical stored envelope. Its
payload is `{ "prpEvent": <canonical PRP event> }`.
Native API tool events omit HTTP request and response bodies from durable history.
This includes serialized provider `contentItems`, legacy HTTP receipts and
`call_api` arguments. Saved receipts retain the operation and call IDs, success
flag, HTTP status and content type when present. The body becomes
`***REDACTED***`, with `apiPayloadRedacted: true` and `apiPayloadSha256`, a SHA-256
digest of the original payload. These digests distinguish duplicate events from
changed payloads; they cannot recover omitted bodies for debugging or replay.
The model receives the original authorized response. Credential-bearing mutation
receipts use encrypted storage separately from this run log.
Checkpoints and completion records apply the same redaction. Completion retries
compare stored and expected payloads through the control-plane port's idempotent
`normalizeReplayPayload` projection, so a saved redacted result can be confirmed
without restoring API bodies. Ports without a projection compare their original
payloads. Repeated redaction preserves the original digest and detects changed
private payloads.
The native runtime passes this projection through backend recovery to the Codex
driver. Recovery compares saved semantic results and turn fingerprints with
provider history through the same representation. A private body change retains
its distinct digest and fails recovery; an unchanged body does not require
restoring its plaintext into the checkpoint.
PostgreSQL JSONB cannot represent NUL (U+0000), which can occur in command
output such as Vite virtual-module paths. The run-event payload column uses a
lossless storage codec for these events: the JSONB projection renders NUL as
the literal `\u0000`, and the reserved `$paperclipRunEventJsonV1` field contains
the original serialized JSON as a doubly escaped string. Ordinary payloads
retain their existing representation. Drizzle reads restore the exact stored
payload before replay, hash validation, redaction, or API presentation. SQL
queries can still inspect ordinary routing fields in the projection; raw SQL
readers of the whole payload must apply `decodeRunEventPayload`. The column
remains JSONB and requires no schema migration.
The writer locks the native `heartbeat_runs` row and allocates the existing
per-run `seq` cursor. A byte-equivalent retry reuses the first row; a changed
retry or source-sequence gap is rejected. Company, issue, agent, run, session,
and runner-source bindings must match the persisted native run. Bootstrap
tickets, reconnect leases, authentication proofs, encryption keys, and raw
credential material are never written to the run log.
These records remain run-log events. They do not create an OpenTelemetry or
Paperclip Telemetry export, and legacy adapters do not use this writer.
## Semantic Settlement Diagnostics and Retry Logs
When native tool arguments fail their authorized JSON Schema, the runner records
`harness.diagnostic` with code `semantic_tool_denied` and `paperclipExecuted: false`.
Its optional `inputSchemaGuidance` retains the same safe repair guidance sent to
the provider: at most three authorized schema locations within 512 bytes.
It includes no rejected arguments, submitted values, or unknown input keys.
The rejected call does not create a semantic input event or execution receipt.
Validated runner diagnostics retain this scalar metadata even when the operation
is an API or case tool. API request and response bodies keep their existing redaction.
If a declared native completion tool is rejected and the turn ends without an
accepted replacement report, the runner emits `session.failed` with code
`semantic_completion_rejected` and `recoverable: false` before the provider's
original turn terminal. It emits no successful result proposal or human review
request. Core uses its existing permanent native failure path and does not start
a missing-result recovery turn. This failure also precedes terminal goal updates;
Core records the original turn terminal before rejecting goal success or further
goal continuation. Recoverable goal failures retain their existing behavior.
The rejection survives same-turn restoration;
a new turn or run binding clears it. A later accepted completion report keeps
its normal authority, and ordinary turns with no completion-tool attempt keep
their existing fallback.
Before interrupting a Codex turn, the runner durably records a
`harness.diagnostic` event with code `provider_interrupt_requested`, the stop
reason, provider turn ID, and pending tool call IDs. A harmless identical result
replay records `semantic_tool_result_duplicate` with warning severity, call ID,
operation ID, and result digest. It does not fail the task or create a user
attention request. A true conflict includes the call and operation IDs and both
result digests in its error; full arguments belong to the canonical semantic
input record, not the error message.
When the tool bridge rejects a recorded result for a call it still holds, for
example a result envelope that names another operation or a value over the
retained limit, the runner still settles that call. It returns a bounded,
non-retryable error to the provider and records
`semantic_tool_result_undeliverable` with warning severity, call ID, operation
ID, the rejection reason, and the rejected result's digest. The provider is not
left waiting for its own client timeout, and the call does not block later
turns or runs.
`run.attach` settles calls that a closed provider turn left pending. The
provider stopped waiting when that turn ended, and the new run's controller
never delivers the old run's results. Each such call is recorded as a
non-retryable failure with an unknown outcome, and the attachment records one
`semantic_tool_result_expired` diagnostic with the call and operation IDs. A
call that a live provider turn still waits on keeps blocking attachment as
`durable_pending_tool_calls`.
Incomplete close emits the bounded `native_session_settlement_incomplete`
runner diagnostic. Its evidence includes runner suspension, provider drain,
final provider state, pending call/operation/source-event IDs and input digests,
and incomplete result-delivery command IDs and statuses. If execution and
cleanup both fail, execution retains its original error identity and cleanup is
attached as `cleanupError`. `runnerRetained: true` means a remote Runner that
could not prove its suspension was left running for the same-run
reattachment instead of being force-stopped; the diagnostic
`remote runner left alive unsettled for same-run reattachment` records it.
Semantic settlement also includes up to 20 content-free failure records, with
the call ID, operation ID, stage (`dispatch` or `persist_result`), and cause.
Causes distinguish an oversized command, a full command journal, known storage
errors, dispatcher rejection, and other persistence failures. Exception messages
and tool results are excluded. These diagnostics do not authorize replay of an
operation whose outcome is unknown.
Instruction writes also commit an `agent.instruction_write_attempted` activity
row and a run-scoped `instructionToolAttempts` entry before permitting the
filesystem effect. They retain the call ID, operation ID, and input digest, not
instruction text. A later transaction rollback cannot erase this attempt proof.
A missing success or definite pre-write failure receipt means the outcome is
unknown, even if the current file contains the requested text. Known validation
and stale-base failures are saved under the attempt's `failure` entry and replay
their original status, message, and details without a new write. Unknown
outcomes remain blocked until reconciled; a
committed receipt with a lost acknowledgement can replay its exact result.
The NDJSON output log appends across repeated `begin` calls for one run. Each
new handle adds an `attemptId` to its lines. If the local file is absent and a
durable S3 mirror exists, `begin` restores that prefix before appending. A
failed restore must not replace the mirror with an empty or partial attempt.
Publishing a restored prefix uses an atomic create-if-absent operation, so a
concurrent restore cannot overwrite lines another attempt has already appended.
Earlier attempts therefore remain available for incident diagnosis. Existing
records without `attemptId` remain readable.
These records use the instance run log and its configured storage. They add no
Paperclip Telemetry or OpenTelemetry export.
## Omitted Unsafe Workspace Export
`workspace_export_omitted` is an informational system event in the local run log.
Its payload is `{ "reason": "restore_unsafe_archive" }`, with `"legacy": true`
when recovering an unsafe failure from an older controller. It records that native
finalization discarded an unsafe export and continued with the accepted result.
It contains no archive names, link targets, or raw error details. It does not
create a task warning, recovery action, Telemetry event, or OpenTelemetry export.
## Native Restart Recovery Run-Log Event
Paperclip writes a `native.recovery.transition` event for native restart
classifications and for graceful restart suspension. A blocked remote retry
records a changed hold reason once; repeated visits with the same hold reason,
an existing future deadline, or a live lease do not append another transition.
Exhausting the automatic admission policy records its terminal disposition even
when the underlying hold reason remains the same.
This immutable run-log
record lets operators reconstruct recovery decisions without exporting data to
Paperclip Telemetry or OpenTelemetry.
The payload contains the restart kind, recovery request id when one exists,
runner disposition, and the controller generation and provider attempt for a
claimed recovery. Live-runner adoption also records the runner PID, process
group, and process-start fingerprint. A non-claim disposition records a bounded
reason instead. Graceful suspension records the signal and confirms that it did
not create a retry run.
The event never includes bootstrap tickets, reconnect leases, authentication
proofs, encryption keys, environment variables, provider credentials, command
arguments, or an unsanitized stderr stream. Detailed failed-attempt diagnostics
remain in the bounded `native_run_finalizations.recovery_history` ledger.
`native.remote_runner_recovery_failed` records the closed `stage` and `reason`
when remote identity verification fails, including admission failures before the
execution coordinator opens. Unavailable probes are retried within a shared
30-second deadline; identity conflicts are not retried. The event distinguishes
an unreadable identity from a confirmed missing process and contains no command
output or credentials. Saved-state read failures use `state_read_failed`; malformed
saved JSON uses `state_invalid`. It does not itself authorize replacement or cleanup.
For `process_not_running`, the event also records `runnerLifecycle` and
`runnerSettled` from the exited runner's saved state. `runnerLifecycle` is a known
runner lifecycle, `other` for an unknown value, or `unavailable` when the state has
none. `runnerSettled` is true only when the outbox is empty and no terminal delivery,
provider cleanup or warm transition is pending. These fields explain how the runner
stopped. The event is not recorded when a scheduled retry resumes a runner that
exited after a settled suspension; see `doc/execution-semantics.md`.
`native.interrupted_turn_outcome_unknown` (level `warn`) is recorded once on a
failed run whose reattachment proved the Runner process gone
(`process_verification` / `process_not_running`) while its last provider turn
had no observed settlement, just before the server requests the automatic
Retry. The payload holds the provider `turnId` (or `null`), the last turn event
type, `replayed: false`, and the closed `runnerExit` fields `stage`, `reason`
and `runnerLifecycle`. It names a turn whose outcome is unknown; the Retry's
fresh session never replays it.
For a prepared remote workspace interrupted before its first executor claim,
that ledger records `remote_execution_never_claimed` with a
`bootstrap_incomplete` disposition. The provider attempt remains zero during
classification; the executor increments it before starting provider work.
A claimed remote bootstrap (the sandbox runner was launched and may have
connected) records `remote_bootstrap_proven_incomplete` with the same
disposition. Recovery admits it only after the previous controller lease has
expired and the Postgres controller checkpoint holds just `run.prepare` and an
unfinished `session.open`, with no committed or acknowledged event. The
executor restores that checkpoint and checks the same proof before it
bootstraps again. After three provider attempts it records
`execution_recovery_budget_exhausted` instead.
A claimed remote run whose previous controller is gone, with no
`native.process_start_requested`, process identity, adapter invocation,
runner-sourced event or controller checkpoint, records
`remote_runner_never_launched` with the same disposition and budget. The
executor records `native.process_start_requested` before a backend may spawn a
runner, so its absence (with no checkpoint) proves nothing reached a runner.
A remote run that a graceful drain suspended just after its runner opened the
provider session records `remote_session_opened_without_turn` with the same
disposition. Its Postgres controller record must hold only `run.prepare` and
`session.open`, with every committed event a session-open event and no event
missing. Before that same-run bootstrap relaunches in the reused sandbox, the
executor stops the interrupted runner it can verify there, so the new runner
does not fail with `runner_ingress_bind_conflict`.
The same rule covers a remote run that its executor already settled as a
same-run `bootstrap_retry` (for example when a shutdown closed the local
runtime listener during startup) and that is due when the server restarts.
For a local Codex, OpenCode, or ACPX turn lost while runnerd was stopped, the canonical `turn.failed`
event retains `error.code: "provider_turn_lost_on_restore"` and
`error.recoverable: true`. It records the interrupted turn, not a final task
outcome. There is no synthesized result for that interruption. An admitted
same-session continuation emits its own `turn.submitted`, accepted turn ID, and
terminal events; only the final outcome closes the run. Reconstructed thread
history retains the original error so reconciliation and live delivery agree.
An ACPX session records this only after the original provider lifetime loses
its fence quorum and the same conversation is restored. Otherwise its
`turn.failed` carries `error.code: "acpx_active_turn_recovery_closed"` with
`error.recoverable: false`, and the run closes with an honest restart message.
After an ACPX restore, the controller can still deliver results for semantic
calls of the lost turn. Runnerd acknowledges each one as `retired` without
sending it to the restored provider, so the calls settle and live semantic
dispatch resumes for the continuation. A restored ACPX provider is a new process
whose MCP request ids restart, so its semantic `callId` values carry a
per-process prefix (`r<hex>:<id>`). They never reuse an earlier call's id in
the same run.
## Native Local Process Stop Evidence
The server writes `native.local_process_stopped` in the same transaction that
clears a local run's process identity, after it verifies that its PID and process
group are absent. The payload contains only those process IDs. Remote process
IDs are never checked against the control-plane host.
A crash can land after `native.process_start_requested` and before the runner's
process identity is recorded. Continuation admission then writes
`native.local_process_stopped` with null process IDs and
`basis: "unidentified_runner_absent"`. It does this only on the host named in
the launch event's `hostname` payload, after it checks two things. First, no
process carries the run's `--run-id` runner argument. Second, no
saved runner control-plane state (canonical or quarantined) belongs to the run;
a provider starts only after its runner connects and writes that state.
The server writes `native.process_start_requested` before a backend can spawn,
and `native.process_identity_recorded` when it stores a new native process
identity. Either invalidates an earlier local stop receipt, including a crash
before the new PID callback. Continuation admission accepts only the latest
server-authored event among these three types; provider
source events cannot supply stop authority. These records stay in the local run
log and do not add Telemetry or OpenTelemetry data.
The server writes `native.prior_owner_fence_refused` when its runner-state
check refuses a run before any launch, because the conversation's saved runner
state belongs to an earlier run that cannot be proven finished. The payload
holds only the closed refusal `reason` (for example
`prior_owner_terminal_state_indeterminate`). Continuation admission reads it,
with no launch, identity, `adapter.invoke` or runner-sourced event, as proof
that the failed run never had a runner, so Retry and new input need no stop
record for it.
Explicit continuation writes `native.stopped_conversation_verified` after it
verifies the retained worker and provider processes stopped. Its payload binds
the requesting `runId`, retained `ownerRunId`, native session, runner instance,
process IDs, retained-state fingerprint, successor run, Retry receipt, and local
workspace identity. The owner can be a preceding terminal
run when the requesting attempt failed before launch with the same process
identity. The server checks host-bound process birth and independent provider
receipts before retiring cleanup ownership. The event does not settle uncertain
actions or authorize replay of the old session.
## Verified Local Codex Replacement Evidence
The server writes `native.stopped_text_turn_verified` in the same transaction
that schedules a fresh successor for a stopped local Codex run. It records the
runner and provider process identities, retained-state digests, provider thread
and turn IDs, and IDs of exactly receipted task-completion calls. The server first
checks the complete turn inventory, process-stop receipt, and execution binding.
Unknown actions or changed retained state prevent this event and replacement.
The record documents why the old execution can be retired. It does not make the
old session resumable, rewrite provider files, or authorize replay on its own.
It remains in the local run log and adds no Telemetry or OpenTelemetry export.
## Sandbox Startup Run-Log Event
Paperclip writes one `run.startup.step` event to the run log for each bring-up
step. This event is a run-log record, not a first-party telemetry event. The
generated telemetry contract does not cover it, so this section is its canonical
contract.
The event payload carries only three fields.
| Field | Type | Meaning |
| --- | --- | --- |
| `step` | string | The bring-up step name, for example `stage.sync`. |
| `durationMs` | number | The wall time of the step. A skipped step reports `0`. |
| `outcome` | string | The step outcome (`ok`, `skipped`, or `failed`). |
The event no longer carries the per-step round-trip count or the provider
duration fields. It dropped `roundTrips`, `providerExecMs`, `providerGetMs`,
`createRuntimeMs`, and `ensureSessionMs`. The startup spans in
[`doc/observability.md`](observability.md) carry that detail now. The
`sandbox.exec` child spans hold the round-trip and provider durations. The
`acp.handshake` step span holds the create-runtime and ensure-session
sub-times.
To read the detailed timing, use the startup spans. The spans need an OTLP
endpoint. A run with no endpoint keeps only the three run-log fields above.
## Run Phase Timing Run-Log Event
Paperclip writes one `run.phase.timing` event to the run log for each
run-lifecycle phase. This event is a run-log record, not a first-party telemetry
event. The generated telemetry contract does not cover it, so this section is its
canonical contract. The producer is `emitRunPhaseTiming` in
`packages/adapter-utils/src/acpx-engine/startup-timing.ts`.
The event payload carries only three fields.
| Field | Type | Meaning |
| --- | --- | --- |
| `phase` | string | The run-lifecycle phase name from the closed allowlist below. |
| `durationMs` | number | The wall time of the phase. A negative or a non-finite value clamps to `0`. |
| `outcome` | string | The phase outcome (`ok` or `failed`). |
The `phase` field is one member of a closed, low-cardinality allowlist. The
producer drops any event whose phase name is outside this allowlist, so a
free-form label never reaches the run log. The allowlist has twelve phase names.
| Phase | Meaning |
| --- | --- |
| `place_workspace` | Place the run workspace. |
| `start_transport` | Start the agent transport. |
| `create_runtime` | Create the agent runtime. |
| `ensure_session` | Ensure the agent session exists. |
| `configure_session` | Configure the agent session. |
| `prepare_turn` | Prepare the turn. |
| `turn` | Run the turn. |
| `end_session` | End the agent session. |
| `settle_reuse` | Settle the session for reuse. |
| `stop_transport` | Stop the agent transport. |
| `sync_back` | Sync the workspace back. |
| `release_staging_lease` | Release the staging lease. |
The payload never carries a command, an argument, a path, an environment value,
or a raw identifier. The event rides the `ctx.onEvent` run-event bridge and is
run-log-only. It needs no OTLP endpoint.
## Workspace Restore and Settled Run Measurements
The heartbeat controller writes these records with `eventType: "lifecycle"`,
`stream: "system"`, and `level: "info"`. The event name below is its `message`
and its payload's `phase`. They remain in the instance run log and add no
Paperclip Telemetry or OpenTelemetry export.
| Event | Payload fields besides `phase` | Meaning |
| --- | --- | --- |
| `workspace_checkpoint` | `mode`, `scannedEntries`, `changedFiles`, `payloadBytes` | Metrics for a validated sparse checkpoint, or a full restore fallback. |
| `workspace.restore` | `durationMs` | Wall time around native workspace restoration. |
| `task.run.measured` | `durationMs`, `outcome`, `durableSettlement` | Wall time through terminal owner cleanup, with the persisted terminal run status and `durableSettlement: true`. |
For `workspace_checkpoint`, `mode` is `"sparse"` or `"full"`. Sparse metrics
count snapshot entries in `scannedEntries`, regular files in the validated
downloaded payload in `changedFiles`, and their total bytes in `payloadBytes`.
Deleted entries and symlinks do not add to the regular-file count or byte total.
Full fallback reports all three counts as `null`: unavailable, not zero. A run
can emit multiple checkpoint records while restoring its repositories. These
payloads contain no paths, filenames, file contents, or credentials.
`workspace.restore` is emitted when the timed restoration exits, including an
exception. It has no outcome field and does not by itself prove restore success.
Its producer is the native workspace sync `onPhase` callback.
`task.run.measured` is emitted only after the controller observes a terminal run
with no remaining native runner ownership. Its duration starts at `startedAt`,
or `createdAt` if the run never started, and includes preparation, execution,
finalization, and cleanup. It is not model-only time; the creation-time fallback
also includes time spent queued. `outcome` is the actual terminal status, such as
`succeeded`, `failed`, or `cancelled`. Provider turn completion can precede this
settlement record. The record does not imply exactly-once emission across
controller recovery, and failure to write it is logged without changing the
already-settled outcome.
## Storage outage waits
A run waits out a file-storage outage on its critical path: when it starts its
run log, restores its saved workspace, or saves it. While it waits, the run's live
status says so, with phase `waiting_for_storage`, for example "Waiting for file
storage before starting the run log (45s so far)". That status lives in server
memory and on the live-event stream only. It writes nothing to the run row or
the run log during the wait.
When the wait ends, the controller writes one record with `eventType:
"lifecycle"`, `stream: "system"`, and `level: "warn"`. Its message is "Waited Ns
for file storage before <step>." or, when storage stayed unavailable, "Waited Ns
for file storage before <step>; it was still unavailable." Its payload holds
`reason: "storage_outage_wait"`, `purpose` (the step), `waitedMs`, and
`recovered`. The record stays in the instance run log.
When the outage outlasts that wait, the run fails. Only a run recorded with
`errorCode: "storage_unavailable"` gets a storage wait: a bounded retry
(`retryReason: "storage_unavailable_wait"`) that starts again after storage is
back. Its error says "Paperclip will retry automatically when storage is back"
only while that retry is scheduled. When the retry budget is spent, the error
says so and asks for Retry. When no storage wait follows, the error says
"Paperclip is not retrying it automatically. Retry or send a message to
continue." That happens when a native recovery already holds the run under
another code (for example `native_continuation_requires_reconciliation`, where
its files are kept and its unfinished turn is not replayed), or when the retry
is refused or cannot be written. A run saved by an earlier release with the
retry promise under another code reads with the same Retry text, with no
migration.
## Durable storage timing
A saved remote workspace is one encrypted archive in durable object storage.
Restoring it before a turn, and saving it after one, runs these steps as named
run phases:
| Phase | Step |
| --- | --- |
| `storage.snapshot.download` | Get the encrypted archive from storage. |
| `storage.snapshot.decrypt` | Hash the downloaded archive and decrypt it to private scratch, at the same time on worker threads. |
| `storage.snapshot.verify_plaintext` | Hash the decrypted archive against its sealed digest while copying it beside its destination, on a worker thread. |
| `storage.snapshot.publish` | Link the verified copy to its destination. |
| `workspace.snapshot.sandbox_restore` | Upload and unpack the archive in the sandbox. |
| `workspace.snapshot.sandbox_upload` | Inside `sandbox_restore`: copy the archive into the sandbox. |
| `workspace.snapshot.sandbox_extract` | Inside `sandbox_restore`: verify and unpack the archive, then publish the restored folder. |
| `storage.snapshot.prior_attempts` | Check whether an earlier upload attempt already landed. |
| `storage.snapshot.encrypt` | Hash the archive, encrypt it and hash the encrypted file, in one pass on a worker thread. |
| `storage.snapshot.digest` | Record the digest of the encrypted file. The `encrypt` pass computes it, so this step takes no time. |
| `storage.snapshot.put` | Upload the encrypted archive (multipart above 16 MiB). |
| `storage.snapshot.verify` | Confirm the stored object (its size when the backend checked every part, otherwise a full read-back). |
A restore runs inside run preparation, so each of its steps of 50 ms or more is
a `run.performance.span` row of the run. A save runs after the turn. Its step
durations are recorded on its `remote_workspace_snapshot` workspace operation,
in `metadata.saveTiming.storage`:
- `priorAttemptCheckMs`, `encryptMs`, `digestMs`, `putMs` and `verifyMs`;
- `bytes`;
- `requests`, the storage requests the save sent. These are `requests`,
`attempts` (SDK attempts across them), `failedRequests`, `retryDelayMs` (SDK
waits between its own retries), `parts`, `throttled` (part sends the backend
rate-limited with 503, 429 or SlowDown) and `backoffMs` (the uploader's waits
before resending a part).
A restore also records its step durations, `bytes` and `requests` on its
`remote_workspace_restore` workspace operation, in `metadata.restoreTiming`
(schema `paperclip.remote-workspace-restore-timing.v1`). A save that reuses an
earlier upload still reads it back; that read-back is counted in `verifyMs`.
A throttled part send counts in `throttled` even when it ends the retries,
and a multipart abort sent during cleanup counts in `requests`.
These are counts and durations only: no object keys, paths or provider text.
## System notices and final replies
Automatic comments with a stored `system_notice` presentation, including
workspace-ready notices, remain available as progress information. They do not
occupy the run's final reply slot. Final response selection still preserves
ordinary explicit agent comments and excludes receipt-bound generated progress.
## ACP terminal failure diagnostics
The shared ACP adapter engine preserves typed terminal session failures in the
run error, the `acpx.error` transcript record, and
`heartbeat_runs.result_json.terminalSessionFailure`. The structured diagnostic
contains the provider category, title, and details. It works when raw provider
tracing is disabled. The existing UI and CLI render the diagnostic as an error,
not as assistant output or an automatic task response.
Issue continuation summaries and session-compaction handoffs retain only the
generic failure category; provider diagnostic prose is not copied into prompts.
Both pinned ACPX patches pass complete title and detail strings to the in-memory
callback. The engine redacts configured environment values (including resolved
secrets with arbitrary variable names), launch environment values outside a
closed allowlist of public process settings, known boolean flags, and run identifiers,
connection URL passwords, the run API key, and common credential forms before
truncation. It removes control characters,
retains line breaks for JSON and stack traces, and preserves up to 4,096 title
characters and 24,576 detail characters. These bounds also keep the escaped
transcript JSON below the server's 64 KiB chunk limit. Longer fields end with an explicit
omission count and appear in `truncatedFields`. The error message includes the
same sanitized text. Ordinary run retrieval preserves the bounded structured
diagnostic even when multibyte text or other result fields exceed the result
byte budget. In that reduced response, title and details have 1 KiB and 8 KiB
byte budgets, including truncation markers. `retrievalTruncated` directs callers
to the full adapter-bounded text in the run error or transcript. Other provider
metadata and action payloads are not copied.
When an ACP request fails with a JSON-RPC error, the engine also keeps the
provider explanation. It reads a string `data`, or the string `data.details`, or
the string `data.message`, in that order. It adds the explanation on a new line
after the error message unless the message already contains it. The combined
text passes through the same redaction and the 4,096-character title bound
before it becomes the run error, the summary, and the `acpx.error` `message`.
A numeric JSON-RPC error code appears as the optional integer `rpcCode` in
`errorMeta` and in the `acpx.error` record. Errors without these fields keep
their existing message and omit `rpcCode`.
Recovery still uses the typed failure category and the adapter's existing
classifier. Provider warnings do not become failures, and timeouts or lost
control channels keep their authoritative failure messages. These diagnostics
stay in the instance's run records and configured run-log storage. They add no
Paperclip Telemetry or OpenTelemetry export.
## Related instrumentation
The sandbox duplex transport also writes one run-log event as one of its three
sinks. See the
[Sandbox Duplex Transport Instrumentation](observability.md#sandbox-duplex-transport-instrumentation)
section in the Observability contract.
## Execution recovery
Cancelled runs retain `resultJson.cancellation`: a closed `source` label
(`operator`, `queued_message`, `shutdown`, `provider`, `transport`,
`control_plane`, `controller_lease`, or `unknown`), whether the stop was expected,
the initiator, reason, and recording time. Recorded stop intent survives adapter
completion. `controller_lease` means the host stopped its own legacy adapter
after it lost the durable controller lease; the reason names a closed cause
(`renewal_rejected`, `renewal_deadline`, or `renewal_error`). The watchdog
retries a failed renewal until the lease deadline, so `renewal_error` means the
attempts still failed with an error when the deadline passed.
All legacy controllers in one server process renew through a single queue on
the private lease connection pool. Renewals that arrive while one is in flight
share the next statement, so a burst of runs costs one renewal statement per
round instead of one transaction per run, and a renewal waits behind at most
one other. A run whose row another transaction holds is skipped and retried
shortly, without delaying the other runs. A renewal only moves the lease
deadline later, so it does not wake live-run recovery; recovery rereads the
moved deadline when the earlier one comes due.
A lease lost during run setup (for example a rejected renewal while preparing
the workspace) settles the run `failed` with error code `controller_lease_lost`,
the same `controller_lease` cancellation evidence, and
`executionRecovery.providerWorkStarted: false`. Recovery retries it as transient
infrastructure instead of blocking on `setup_failed`. A setup-phase error after
provider dispatch began keeps `setup_failed`, so started work is never retried.
The local lifecycle event includes this evidence. Started cancellations without
an expected stop are also reported to Sentry; its cancellation diagnostics contain
only source, expectedness, initiator type, and the closed lease-loss cause, never
initiator IDs or reason text.
A lease loss also records `resultJson.controllerLeaseLoss`, including a loss
during setup that fails the run before adapter dispatch. It holds the closed
cause and the watchdog timings in milliseconds: `sinceRenewalMs` (since the
last successful renewal), `renewalPendingMs` (age of a renewal operation still
in progress, including its retries and retry waits, or `null`), `deadlineLateMs`
(how late the deadline callback ran, when the lease deadline ended the run,
otherwise `null`), and `renewalCount`. A long `renewalPendingMs` can mean a
stalled database query or repeated failed attempts with retry waits; the cause
label tells them apart (`renewal_deadline` or `renewal_error`). A long
`deadlineLateMs` points at a blocked event loop. Failure reports carry these
bounded numbers and the cause label only.
Historical ambiguous cancellations stay `unknown` and do not authorize replay.
Provider tool-definition validation failures use
`provider_tool_definition_invalid` / `configuration`. Automatic retry and
continuation recovery stop until the configuration is repaired. Classification
uses raw provider diagnostics in memory before redaction; stored diagnostics
remain redacted and bounded.
Provider identity diagnostics remain in the local run log. They record the notification method, expected and received thread/turn identifiers, and the classification (root, verified descendant, stale, unrelated informational, or invalid authoritative). They omit the original provider payload and credentials. Repeated informational notices are bounded.
Ignored unrelated Codex notifications use `harness.diagnostic` with code
`codex_unrelated_information`. The payload retains only the bounded provider
method and expected/received thread and turn identifiers. Account updates, skill
changes, and unrelated thread information do not create a provider notice in
chat. Chat also omits the matching notice stored by older runners. Real provider
warnings and errors remain visible.
Provider rate-limit windows use `provider.notice.recorded` with category
`provider.rate_limits` and scope `account`. Codex sends them as
`account/rateLimits/updated`. Claude sends them as `usage_update` metadata
(`_claude/rateLimit`) when it uses a subscription sign-in. The runner merges
sparse Codex updates and records a notice only when the windows change. The
notice has one `rateLimits` detail: a JSON snapshot with schema
`paperclip.provider.rate_limits.v1` (windows with used percent and reset time,
credits, and the reached window). Chat does not show the notice. The server
keeps the latest snapshot in the agent's runtime state under
`providerRateLimits`, and the task properties pane shows it as the assignee's
**Usage limit**. The snapshot is display data. It does not hold or start work;
the provider-limit wait reads typed run failures.
Harness status notices use `provider.notice.recorded` with a status category.
Codex `error` notifications read `error.message`. A notification with
`willRetry: true` becomes a recoverable `provider_retry` notice. A final error
keeps category `error`. The `CodexErrorInfo` kind, the HTTP status, and the
reconnect attempt count become notice details. Claude compaction, model
fallback, and SDK informational banners arrive as ACP message text without a
message identity. The ACPX layer records them as `context_compaction`,
`model_fallback`, or `provider_informational` notices instead of answer text.
Compaction start and end share one notice identity and carry a `Phase` detail.
The task chat shows these notices as inline status chips. A retry chip fades
after later provider work. A final provider error shows the real provider
message. Generic warnings stay out of chat as before.
Recovery lifecycle events retain the original structured failure code, retry attempt, next retry time, and predecessor/successor identifiers. Durable status delivery uses an idempotency marker; delivery grants no provider authority. Failed publication is retried without repeating provider work. These records are not first-party Telemetry.
An unstarted legacy conversation retry waiting for execution cleanup retains
`resultJson.executionWait` with the issue, blocking run, and
`execution_owner_active` cause. A local lifecycle event records the wait and the
unchanged scheduled retry attempt once per blocking run. Repeated scheduler
checks update the same retry row without repeating that event. Promotion removes
this ownership-wait marker after cleanup; other execution holds retain their
existing dispatch gates.
If execution-continuation setup finds that a task no longer exists, is closed,
or its owner changed, the existing cancellation settlement records
`continuation_task_ownership_changed`. The run and wake request become cancelled
before adapter dispatch, with the run-log message
`stale execution continuation cancelled before dispatch`. Immediate recovery is
suppressed. A comment the cancelled run carried that no run read is saved
first when an agent run closed the task: a person's comment, or another agent's
comment posted with `resume: true`. The release then reopens the task and
delivers that comment once. Another agent's comment without `resume: true`
stays communicative and does not reopen the task. Missing source context, authorization failures, and other setup
errors retain their failure classification. An untyped error with the same
message is also still a failure; cancellation requires the typed ownership guard.
Bounded retry exhaustion writes one lifecycle receipt per run, retry reason,
scheduled attempt, and retry limit. Repeated or concurrent recovery checks reuse
that receipt, including receipts from earlier builds, without advancing the event
sequence or publishing another live event. Attention reads select the latest
matching receipt in PostgreSQL and project only the run's issue/task identifiers
from its context, so historical duplicate receipts cannot multiply run contexts
in server memory. Existing duplicate events do not require deletion or migration.
### Workspace restore failures
Legacy adapter results can carry `workspaceRestoreFailure` with the code
`restore_permission_denied`, `restore_lock_timeout`, `restore_unsafe_archive`,
or `restore_failed`. The heartbeat records `workspace_restore_failed` and keeps
the run failed and the workspace-finalization barrier closed. Available output,
usage, session metadata, and the previous execution outcome survive settlement.
`executionBeforeRestore` retains the earlier error code, exit code, signal, and
timeout flag. The ordinary redacted error field retains an earlier error message.
The chat reports the restore phase separately from a missing final response.
A saved-plan link requires a stored document and its run-bound revision or a
matching run-bound review record. Older unclassified failures use neutral wording. Diagnostics show only
a validated relative member path, never an archive link target or host path.
An unsafe archive or an outbound confinement refusal keeps the existing execution recovery hold, including across
conversation resets. It cannot start another model turn until an operator uses
the existing recovery action to record `executionReconciliation` with
`workspaceRepairEvidence` (20–12000 characters). This evidence must describe
verified safe staging or repair for the referenced failed run. It does not grant
plan approval. Saved comments, document revisions, and confirmation IDs and
states stay unchanged. Recovery uses the existing delivery identity and links
the successor to the original failed run. Repair does not reset the automatic
retry budget. Transient failures retain the existing
bounded retry policy. Archive confinement remains required.
Sandbox restore tasks also write a `Workspace restore diagnostic` line to the
run log on failure. `phase` is `workspace` or `asset`, so a failed staged-asset
copy-back (such as credentials) can be distinguished from workspace restoration. The line
contains only an allowlisted OS/transport `errorCode` (otherwise `unknown`), an
optional numeric HTTP error status, and an optional bounded process exit code.
Up to four nested causes are inspected. Messages, URLs, filesystem paths, asset
names, credentials, and response bodies are excluded. Every failed outbound
task emits its own diagnostic; nested repository failures are logged once by
the enclosing workspace task. The original error and restore safety policy are
unchanged. These lines stay in the instance run log and its configured durable
storage, and are not new first-party telemetry events.
Native sandbox `environmentSyncOut` errors preserve allowlisted error codes and
bounded HTTP/exit statuses across worker RPC for this diagnostic line. The host
revalidates the envelope and keeps the original failure and recovery policy.
Provider messages, paths, response bodies, credentials, and arbitrary error data
are not copied into the new envelope. Older workers can still report `unknown`.
The optional `step` identifies the failed restore operation. For
`phase=workspace` and `step=git_integration`, `gitCommand` identifies one fixed
command family (`rev_parse`, `symbolic_ref`, `merge_base`, `merge_tree`,
`commit_tree`, `update_ref`, or `log`). `gitFailureKind` is `merge_conflict`,
`invalid_object`, `ref_conflict`, `permission_denied`, or `unknown`; it is a
bounded diagnostic clue, not a new recovery or retry decision. Only supported
exit/OS codes and recognized Git messages produce a specific classification.
No command arguments, stderr, filenames, repository URLs, or ref names are saved.
The same closed fields persist in `workspaceRestoreDiagnostic` and are
revalidated before projection into an enabled Sentry failure report.
## Codex resume usage snapshot
The native runner retains a bounded local `harness.diagnostic` event with code
`codex_resume_usage_snapshot`. It identifies `thread/tokenUsage/updated` as
`resume_usage_snapshot`, retains the reported thread and completed-turn IDs,
and records cumulative usage counters. It does not include provider credentials
or message content. The event establishes the accounting baseline; it is not a
new billable usage receipt or a user-facing provider warning. Other provider
identity checks remain in force.
## OpenRouter serving host
OpenRouter can route one model across many hosts, which may serve different
quantizations. OpenCode runs routed through OpenRouter record which host served
each model call. The runner's provider proxy reads the top-level `provider` and
`id` fields from the first streamed chunk, or from the whole JSON body when the
response is not streamed. It forwards every byte to OpenCode unchanged and
before it inspects them, so the first token is not delayed. It scans at most
the first 64 KiB of a stream and 1 MiB of a JSON body. It skips encoded bodies,
non-2xx responses, and any malformed or missing field. A recording failure
never affects the response or the turn.
Each `item.completed` usage event (`payload.kind: "usage"`) carries two fields
in `payload.usage`:
- `upstreamProvider`: the host that served the most recently finished model
call when the usage was reported, such as `Novita` or `DeepInfra`.
- `providerRequestId`: that call's OpenRouter generation ID, such as
`gen-1762981234-AbCdEfGh`.
The proxy records a call's host when that call's response ends, not when the
request starts. OpenCode reports a step's usage only a few milliseconds before
it sends the next model request, so a step's usage still names its own call's
host while the next call is in flight. A call whose host the proxy could not
read records no host when it ends, and a session clears the recorded host when
it attaches another run. Such a call therefore carries neither field. It never
inherits an earlier call's or an earlier run's values.
The runner's internal `usage.reported` event carries the same values, using its
existing `providerRequestId` field and an optional `upstreamProvider` field.
The run's `usageJson.providerRequestId` and its cost event name the first
generation ID the run reported. That value stays the same as later turns
report, so a late usage receipt still matches the charge. Remote Runner runs carry
the same ID in each saved usage receipt's producer event and in the receipt
itself. A receipt saved before this field existed still replays when the
Runner now names an ID. Per-call IDs and
hosts remain in the run log. Runs whose provider reports no host carry neither
field, and `usageJson.providerRequestId` stays `null` as before.
These are run-log fields. They stay in the instance database and are not sent
to Paperclip Telemetry or OpenTelemetry.
## AI subscription contention
A fresh task execution cannot enter this wait. A run that already entered this
wait writes an informational `lifecycle` event to the local run log. Its
payload contains only `retryScheduled`, a boolean that reports
whether the scheduler created a retry.
The message distinguishes an automatic retry from work that is no longer eligible.
This pre-provider wait records `ai_connection_busy` on the cancelled run and does
not consume the provider-failure retry allowance. The event contains no credentials
and creates no Telemetry or OpenTelemetry export.
## Agent directory waits
A run whose instruction preparation finds another run of the same agent using
its files writes one informational `lifecycle` event to the local run log, with
`stream: "system"`. It records no file contents, paths or credentials, and
creates no Telemetry or OpenTelemetry export.
- A resumed native run that stays on its own run writes the message "Another
run of this agent is still using its files; this run continues when it
finishes". Its payload has exactly two fields: `retryReasonCode`, always
`agent_directory_operation_busy`, and `nextAttemptAt`, the ISO time when the
same run is scheduled to try again.
- A run that has not yet started native execution is cancelled with error code
`agent_directory_busy` and writes "Another run of this agent is still using
its files. This task will continue automatically." when an
`agent_directory_busy` retry was scheduled, or "Another run of this agent is
still using its files, and this task could not be rescheduled." when it was
not. Its payload contains only `retryScheduled`, a boolean.
Neither wait consumes the provider-failure retry allowance. Each stops after 30
deferrals, after which the run follows the normal failure path.
## Task history reads
When a native run reads its task's comments through `get_task_history`, the
server writes one `task_history.read` event (`stream: "system"`, level `info`)
to that run's log. Its payload has one field, `comments`, which maps each
returned comment ID to that comment's `updatedAt` timestamp. It contains no comment bodies or
author identifiers. The event creates no Telemetry or OpenTelemetry export.
Per-comment delivery (`deliveredQueuedCommentIds`) reads this event together
with the run's `run.result.accepted` event. A read alone is not delivery: the
run may have read the comment and set it aside. A comment counts as received
only when a run of the task's assignee succeeded after reading the comment's
current revision, and that run's accepted result cites the comment as a
`comment:<id>` evidence reference (top-level `evidence[].ref` or a criterion's
`evidenceRefs`). The comment's queued wake then neither reopens the finished
task nor starts another turn. An edit after the read, a read by a run that did
not succeed, or a read the result does not cite leaves the comment owed, so it
is delivered once in a new turn. That turn's wake marks the comment
`readByEarlierTurn` when an earlier finished run read its current revision, and
the prompt asks the agent to check the current task state before acting so
already-applied work is not repeated. `get_task_history` returns
`arrivedDuringThisTurn` (the `comment:<id>` refs of comments that other authors
posted or edited after the run started) with instructions to cite each applied
comment. The check reads only the assignee's succeeded runs of that task that
finished at or after the comment's current revision, through the existing
run-binding and `(company_id, run_id)` event indexes. A saved follow-up
resumed after a remote Stop uses the same rule. `taskHistoryReadsForRun` lists
the comments one run read, with each read's time and the comment revision it
saw, whatever the run's outcome. A failure to write the event leaves the
comment owed in the same way.
When a run reads a finished child task, the server writes one
`child_task.read` event (`stream: "system"`, level `info`) to that run's log.
The reads that count are `get_task_context` on the parent task, and a
successful agent request with a run token for `GET /api/issues/:id`,
`GET /api/issues/:id/documents` or `GET /api/issues/:id/documents/:key` of a
child in `done`. The route writes the event only after the requested content
has loaded, so a 400 or 404 records nothing. The payload has `parentIssueId`
and `children`, which maps each finished child ID to the revision the run
received: its `status`, `updatedAt` and `statusDecisionId`. It contains no
titles, summaries or document bodies. The event creates no Telemetry or
OpenTelemetry export.
`hasPendingNativeChildCompletion` reads this event. A child-completion wake no
longer keeps the parent's done claim open when the finishing run received that
child in `done` at the completion the wake reports: the received
`statusDecisionId` is the wake's completion decision, or the received
`updatedAt` is at or after that decision was saved. The event's own write time
does not count, so a delayed write of an older revision cannot deliver a newer
completion. The wake is then cancelled at release instead of starting another
parent turn. A read while the child was still open, a read of an older
revision, a read by another run, or a failure to write the event leaves the
child result owed.
## Managed Agent File Save Receipts
The server writes `instruction_save` after managed file collection or a warm
turn checkpoint. The payload includes the save state, instruction entry path,
storage warning, and error code/message. Agent-directory receipts identify
`contract: "agent_files"` and the applied candidate hash. Legacy instruction
receipts instead identify the saved revision.
A validated warm checkpoint reports `saved` or `unchanged`, even though its
working directory remains owned by the live session. An unstable checkpoint
reports `pending_collection` until stopped collection produces a final receipt.
Successful checkpoints can include `checkpointStats`: `scannedEntries`,
`hashedBytes`, `copiedFiles`, and `copiedBytes`. These counts describe that
capture, not cumulative traffic or an atomic snapshot of background writers.
They contain no file contents. The receipt remains in the instance run log;
it adds no Paperclip Telemetry or OpenTelemetry export.
If instruction-copy release throws, `instruction_cleanup` records a warning
with payload `{ "state": "deferred" }`. The existing working-copy recovery sweep
retries cleanup. This event preserves the run outcome and does not claim a file
save; `instruction_save` remains authoritative for collection. The event contains
no raw exception, host path, file contents, or lock-owner metadata. Failure to
write the warning must not replace the provider outcome or stop lease release.
## Agent observability availability
`agent_observability.unavailable` is one warning for a preparation snapshot that
cannot apply company export settings. Its payload contains only a closed reason
and the settings revision. It carries no endpoint, credential, secret reference
or resolver error. Recording the diagnostic never contacts a destination.
## Runtime activity
`runtime.activity.recorded` is a P1 Runner PRP event with a closed
`{ namespace, activityKey }` payload. `namespace` groups a producer's activity
types under a stable name. `activityKey` identifies an activity within that group.
Consumers use the pair to select a response to activity in the current run.
For example, `{ namespace: "example.computer", activityKey: "desktop.used" }`
records successful desktop use. Core can map that activity to a Desktop viewer
for the task's running agent. The authenticated event envelope supplies the run,
session, source sequence and emission time. Core resolves the company, task and
agent through the run's existing authorization checks.
Both activity identifiers contain 1–160 Unicode code points and exclude C0
controls and DEL. Callers cannot supply run, task or agent identity in this payload.
The ordinary authenticated Runner event is the durable record. Normal source-event
acknowledgement and replay deduplicate delivery. Existing run-read authorization
and secret redaction apply to saved and live events. Recording reported activity
is not independent proof that it happened and grants no resource permission.
The native transport forwards runtime activity as a run-scoped fact. It retains
the authenticated run identity across warm session attachment. Provider messages
cannot impersonate this transport-owned notification.
New-run activity waits for warm attachment to finish. A failed attachment does
not retarget queued activity to the previous run.
Runner accepts local activity records only with an opaque context
captured before the operation. Warm run attachment and process recovery rotate
that authority. Runner supplies the run identity from its active authenticated
binding. Optional hint failure does not fail the underlying operation. No output
text is scanned and no synthetic receipt is written.
`GET /api/heartbeat-runs/:runId/events?view=plugin-panels` reads these events and
returns only installed agent task-panel contributions under normal run-read and
telemetry-read authorization. Ordinary PRP reads/live delivery retain run-secret
and current-user redaction. The task page uses one history read on mount/reconnect
and the existing live transcript stream, then makes a bounded plugin readiness
check. No panel discovery timer scans runs or provider state while connected.
## Post-result failure diagnostics
`native.post_result_failure` records `stage: post_result_execution`, a closed
`category` (`io_not_found`, `io_permission_denied`, `io_failure`, or
`execution_failure`), and `resultAlreadyAccepted: true`. It requires an accepted
selected result and exact current, unexpired controller authority. A busy diagnostic
row lock is skipped. Reporting is best effort under the existing 60-second control
bound; the sink can finish later and emit a late event. The `native_session_lease_lost`
control classification stays unchanged. The labels grant no recovery, result,
Stop, provider-work or payment authority and do not establish a hosted failure cause.
## Payload storage
Structured events in this guide remain in `heartbeat_run_events`. When the
instance's existing file storage uses S3, new stdout/stderr NDJSON logs use the
same stack bucket and configured prefix under `run-logs/`, with existing run
references in Postgres. Local-disk instances remain local. Raw provider
traces have their own restricted, encrypted object path. See
[run evidence storage](run-evidence-storage.md) for batching, durability,
configuration, migration, retention, and qualification boundaries.
### Event commit stall diagnostics
A Runner event commit stall records the first safe error code in the error and
up to three nested `cause` objects. This retains database and transport codes
when an adapter wraps the original error. Each code has the existing 80-character
identifier limit. If no safe code exists, the diagnostic uses the outer error's
safe name. It never copies exception messages, SQL, or query arguments. Cyclic or
long cause chains stop at the same four-object limit.
This only improves the recorded cause. Event acknowledgement, replay integrity,
the bounded stall policy, and operator recovery requirements stay unchanged.
A later successful continuation does not establish the cause of an earlier
failure that recorded only `Error`.
### Failure cause of generic run failures
A run that fails with a generic code (`adapter_failed`, `setup_failed`) keeps
that code, because callers compare against it, and records
`resultJson.failureCause`: one identifier of at most 80 characters
(`[A-Za-z0-9_.:-]`). For a thrown error it is the provider's HTTP status
(`provider_http_429`) or category, else the first safe code in the error and up
to three nested `cause` objects (`40001`, `ECONNRESET`, the same extraction as
the Runner's event commit stall diagnostics), else the error's class name. For
an adapter result without a code it is the error family, `signal_<name>`,
`process_cancellation_failed`, `exit_code_<n>` or `adapter_reported_error`. It
never copies message text. The server's failure log line carries the same
cause, and the board's run list shows it, for example "adapter failed (40001)".
Older rows have no cause and read as before.
Run error text keeps SQL statements off the board: a Drizzle `Failed query:`
statement and its parameters, or unwrapped statement text with `$n`
placeholders, is replaced with "Database query failed (<cause>)". The wrapper
text before the statement is kept.
## Chat outcome read receipts
Two run-log events record what the conversation agent saw before it replied.
`chat_handoff.completed_read` records that a run read a task it handed off, in
the done state, with the task ID and status version. `chat_interaction.declined_read`
records that a run listed its conversation's interactions and saw a declined tool
action, with the issue ID, interaction ID and status. Neither carries task text,
tool arguments or results.
A completion or decline continuation that an earlier reply already reported
creates no run. The server checks for a reply by the same agent in the same
conversation, posted after the outcome, from a run that has one of these receipts
for that exact outcome. A completion reply must also link the task. The
completion delivery is then marked delivered with that reply as its response,
and a deferred wake ends `skipped` with the reason "Already reported by an
earlier reply in this conversation".
The conversation's handoff listing (`GET /api/issues/{id}/chat-handoffs`) shows
that reply as `completionReportCommentId` once the latest completion is delivered.