DottaandPaperclip
53f480d1b3
## Thinking Path > - Paperclip manages AI agents and their saved work. > - Product E2E tests must check saved personal files across separate turns. > - The server records file collection in the durable `instruction_save` run-log event. > - The warm test read optional terminal result metadata instead and rejected a recorded successful save. > - This pull request reads the complete event stream for the current run and uses its latest save receipt. > - The prompt now specifies the exact nine bytes required by the temporary-file transfer count. > - Human review actions select the actual interaction card, which can share its interaction ID with a navigation button. > - A later real three-turn run exposed a missing reply after workspace finalization moved to its durable recovery owner. > - Reply recovery handled completed tasks but omitted ordinary tasks that awaited review. > - This pull request also publishes their recorded final reply once while preserving the pending review. ## Linked Issues or Issue Description **What happened?** The [retained live continuity attempt](https://github.com/paperclipai/paperclip/actions/runs/41477302876/job/125420001074) saved the correct memory file and unchanged 8 MiB binary. Its durable run log contained a successful `instruction_save` receipt. The oracle read `resultJson.instructionSave`, which was absent, and failed before turn two. The temporary file also had a trailing LF. Its prompt allowed that result, while the byte-count check required nine bytes. **Expected behavior** The oracle must use the documented authoritative save event. A missing receipt, a later failed save, incorrect transfer counts, or another run's events must fail. The prompt must specify the bytes that its independent check requires. **Steps to reproduce** 1. Run `daytona-warm-continuity.runner-codex.daytona.warm-three-turn` with the configured protected credentials and immutable image. 2. Complete a native turn whose terminal result has no `instructionSave` field. 3. Compare the oracle output with the same run's successful `instruction_save` event. A subsequent [three-turn attempt](https://github.com/paperclipai/paperclip/actions/runs/41495337334/job/125451304553) passed every personal-file assertion and reused the same warm process. Its first final reply was recorded as a completed agent-message event but was absent from comments. The log shows `native_workspace_finalization_busy`; the durable recovery path excluded `needs_review` replies. The final message matcher correctly failed. This failure remains recorded. The observed application revision was `aab87ce70744660a4a21c77278b0ede267c04720`. Related work: #14695 and #18494. The retained failed result keeps its original grade and definition. ## What Changed - Read all run-event pages through the existing bounded collector. - Require events to belong to the requested run and select the latest `instruction_save` event. - Retain strict byte counts and reject missing, failed, pending or superseded receipts. - Specify nine ASCII bytes and no trailing newline for the temporary file. - Scope human review actions to the interaction card and its exact interaction ID. - Advance the warm suite definition to version 5 and document its evidence source. - Recover recorded final replies for ordinary native tasks with an accepted `needs_review` result and terminal state. - Preserve pending reviews, existing reply selection, company/turn/digest checks and external governed-response authorization. - Update the task feature and agent access maps. ## Verification - Real PostgreSQL regression: deferred review reply discovery failed before the production repair (expected one candidate; found zero). The final-message event and pending review were present. - Recovery cases cover one publication, no repeated provider run, unchanged review, and rejection of absent/wrong-turn final messages, bad digest and suppressed comments. - Red: seven new boundary regressions failed against the old receipt lookup. - Green: 169 focused oracle, event pagination and catalog checks passed. Product E2E typecheck passed after rebuilding local package declarations. - Retained-evidence replay confirms that the new lookup finds the real successful save event. The original attempt remains failed because its temporary file included an LF. - Recursive typecheck passed before the final recovery change; server typecheck passed afterward. The recovery and control-plane suites passed all 112 tests. Full build and map checks passed. - The broad local test run reached one unrelated 30-second timeout in `attention-retention-work.test.ts` (maintained-review/confirmation batching). The same unchanged case then passed on one targeted retry (39.9 seconds including module load). The original failure remains retained. The broad run also reported one unchanged workspace-conflict case, which passed on its targeted retry (14.85 seconds including module load). Both original failures remain retained. The redundant broad local run was stopped after current-head CI completed all required gates: 55 successful checks and 3 selection skips. No all-green local broad-run claim is made. - The separate artifact-transfer cell already passed on the recorded application revision: independent source and delivered bytes matched at 407,485 bytes, with a completed task, succeeded native run and cleared execution locks. This does not establish a three-turn continuity pass. - [Real native three-turn acceptance](https://github.com/paperclipai/paperclip/actions/runs/41499102994/job/125465795210) **passed** at `3b207a42943e9eeb371d4f2d5623de72093eb6d5`, suite definition 5. Turn one exercised `committed_task_response_recovered`. All three completed final replies were persisted once and in order. All three strict personal-file receipts passed. The unchanged 8 MiB binary survived; later transfers copied only 36 and 54 bytes. The lease was created, resumed, and resumed with the same runner PID/start time, runner instance, native session and provider session. Final task was `done`, all three runs succeeded, and checkout/execution locks were null. - The aggregate startup summary has no separate provider-process PID proof. The raw run records independently prove the stable runner process and provider session stated above. This single native cell does not qualify the legacy cell or a full provider matrix. - The passing campaign reused the signed immutable runner image. It required no disk expansion. Earlier campaign image builds used the recorded temporary 120 GiB mitigation; default 60 GiB build capacity remains unqualified. All scoped completed probes were removed and independently returned 404; org CPU/disk usage returned to zero. This does not resolve unknown historical resources under a different credential identity. - Current-head Greptile is 5/5; review-thread pagination is complete with zero threads. All 55 current-head checks are successful; 3 selection checks are skipped. The paid campaign, result enforcement and report publication all completed successfully. ## Risks - Publishing a reply must not approve the work. Recovery only copies the accepted turn’s completed final message. It preserves the pending review and does not start an executor. - Conversation and external governed-response paths retain their existing separate authorization. The latest save event can record failure after an earlier success. The oracle rejects that case. Event pagination, missing receipts and cross-run evidence also fail closed. The exact temporary-file byte count remains required. The review selector still requires the intended interaction ID and enabled action on its actual card. The production change is limited to ordinary internal-task reply recovery. It still requires an accepted result, a matching successful terminal disposition, the completed final message, and existing publication authority. ## Model Used OpenAI Codex, based on GPT-6, with shell and code tools. The exact deployment variant and context-window size are not exposed in this session. ## 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 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 - [x] All Paperclip CI gates are green - [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups - [x] I will address all Greptile and reviewer comments before requesting merge --------- Co-authored-by: Paperclip <noreply@paperclip.ing>
Paperclip agent access map
This map accompanies the feature map. It records APIs, agent entry points, permission boundaries and verification for the same capability IDs. The shared contract defines the format and maintenance rules.
The chat scenario catalog uses these same IDs for human requests and expected outcomes. Its scenarios retain the access boundaries here; neither a CEO role nor a planned scenario grants board authority or establishes a working production tool binding.
Use full capability IDs when referring to access contracts outside their file:
onboarding/company means the company capability in onboarding.md in both maps.
The feature index lists feature and
capability IDs; IDs and references
explains entry-point references and stable naming.
Browse access contracts
The same-named feature recipes define navigation and capability IDs. Run
node scripts/storybook-catalog.mjs for the generated
maps/generated/agent-access-map.md index, or browse the
published feature guide.
Add or update the affected companion; there is no central registration or count.
Existing baseline source references describe the original inspection, not current verification. Record new executed evidence in the PR, issue, or CI artifact with its revision and scope. Do not refresh unrelated headers as bookkeeping.
Using an entry
Read its Policy and Gaps before selecting an agent path. Native fallback, legacy
REST/CLI, stdio MCP and a personal assistant connection can have different
authority and availability. Use a real admitted production binding; a scenario
declaration is not an installed tool. A restricted entry records current
enforcement and leaves any unresolved delegation decision visible.
For API catalog disagreements, follow the route/service authorization and record
discovery drift. For not-applicable, retain the concrete exception and local
path. Split shared business operations from client-only presentation.
Use maintain-verification-skill to reconcile changed source and both maps.
Use verify-paperclip and existing API/runner/MCP fixtures to collect the
permitted-agent and denial evidence; report blocked/not-run coverage honestly.
Run node scripts/check-maps.mjs after editing. The structural check does not
certify permissions, API catalog coverage, tool admission or live behavior.
Native REST transport
The call_api parity change reconciles affected baselines against
58e16dbddfc9c8d0cf60867f9d9847f0feda055b and the changes in this PR. Ordinary
REST operations use the same route authorization as legacy agent REST. Dedicated
tools remain optional for those operations; protocol endpoints still need their
existing clients. Board-only routes remain restricted for agent identity.
Native search_api / call_api is enabled by default, but explicit server,
company-allowlist and binding settings can narrow admission. It requires a running
authorized native run, current task assignment and execution-run binding, company
and budget authority, task privacy, and a permitted work mode. Ask and Plan retain
mutation restrictions, and runs retain their admitted JWT scope during restart
recovery even when task mode changes. A waiting
agent does not gain independent native control just because a REST route exists.
Connection-management reads support Standard, Ask and Plan; mutations require
Standard. Skill-test runs and nonstandard task-bridge keys cannot administer
connections. These gates supplement each operation's grants, resolved resource
and target scope, and responsible-user authority. See the
rollout policy
and connection authorization.
Credential-bearing replies, including company exports and their legacy alias, stay inline under a 10 MiB limit and use encrypted mutation receipts. Other saved API snapshots are limited to the issuing actor and responsible user, plus authorized administrators. Artifact uploads enforce the same access checks as downloads; private snapshots are excluded from company exports and public company logos. Explicit deliverables use the existing sharing path. Native event history, checkpoints and completion records omit API bodies while preserving receipt metadata and replay digests. Restored webhook secrets use the same encrypted receipt and inline response policy as newly created webhook secrets. Task image access requires the company boundary and the task permission; skill-test and low-trust agents do not need company-wide browsing authority. Completion replay applies the same idempotent projection to stored and expected payloads; the run-log guide describes its fields.
Sources: API client,
catalog,
native authority,
credential storage classification,
snapshot access,
and history redaction. The affected entries remain
baseline and source-only; this reconciliation does not certify every operation.
Targeted real-route proof lives in the
API integration suite,
including owned routines, skill tests, scoped secrets, proposals and private export replay.