Files
DottaandPaperclip 992f720262 fix: make runner task context ownership explicit (#13753)
<!-- Write all pull request text in Simplified Technical English
(ASD-STE100). -->

## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work.
> - Task descriptions, comments, continuation data, skills, and
execution rules enter several agent adapters.
> - The same source can be rendered by more than one automatic input
carrier.
> - Failed resumes can also rebuild input from stale or compact context.
> - This pull request gives each Paperclip-owned source one delivery
owner and preserves the required transport boundaries.
> - It adds deterministic adapter, interaction, runner, and browser
tests for these boundaries.
> - The benefit is more predictable context delivery with explicit
evidence for later live qualification.

## Linked Issues or Issue Description

Related: #13144 removes a duplicate environment payload and bounds wake
lists. Related: #11360 addresses Hermes resume behavior. This pull
request preserves compatible active-session formats while repairing
context ownership and stale question creation.

**What happened?**

Task descriptions and comments could enter more than one automatic
context block. Native transports could wrap a complete model input in a
second task envelope. Some legacy and gateway adapters could omit the
owned assignment on ordinary tasks or rebuild a failed resume with stale
compact context. A continuation could also request a question after
newer human comments had arrived.

**Expected behavior**

Each task or comment source has one automatic model-facing owner.
Distinct comment IDs and repeated wording remain distinct. Fresh
fallback attempts rebuild the required full context. A question request
is rejected when newer queued human direction makes it stale. Harness
access policy remains owned by execution configuration.

**Steps to reproduce**

1. Build a task with a description and current comments.
2. Capture the actual adapter or runner input.
3. Compare source ownership and task-envelope nesting.
4. Queue a human comment before a continuation requests a question.
5. Trigger a failed resume and inspect the fresh retry input.
6. Run the focused adapter, interaction, runner, and browser checks.

## What Changed

- Add shared prompt-section selection at the provider-attempt boundary.
- Deliver owned assignment context through native, legacy CLI, ACP,
gateway, cloud, Pi, Kimi, Grok, Gemini, OpenCode, Cursor, OpenClaw, and
Hermes paths.
- Rebuild full or compact context after resume recovery changes the
attempt. Add native and Claude ACP tests of actual recovery requests.
- Preserve custom templates, loaded instruction files, execution
policies, and older active-session formats.
- Record continuation source metadata and reject stale question creation
under the issue-row lock.
- Add explicit Product E2E context-integrity profiles, prerequisite
gates, credential-isolation checks, and report fixtures.
- Bypass service-worker forwarding for same-origin Vite development
modules. A real Chromium test fails with resource exhaustion before the
repair and passes after it. Production asset caching keeps its existing
policy.
- Add browser diagnostics and service-worker module-loading regressions.
- Add an explicit zero-retry eval option. The default retry behavior
remains unchanged. Each campaign records its effective policy.
- Remove the model-facing working-directory sentence from four prompt
builders. Existing workspace, sandbox, permission, and custom-template
configuration remains unchanged.
- Align the everyday workflow assertion with the current 47-entry
catalog.

Compared with current upstream master, the branch carries the
context-ownership implementation and its tests, the explicit
context-integrity catalog and evidence harness, and the focused browser
regression checks.

## Verification

**Merge assessment:** focused regression evidence supports merge. This
is not full completion of the original broad qualification matrix. The
maintainer has authorized merge after fresh verification of the master
integration.

- Current head: `bbd52f82114eabf09bc7b1a7e97d54a5b43bbc00`. This
integrates current master `2f585ef26a1814fa209715242d1ca791b63e4c4e`.
All 14 conflicts are resolved. Cancellation checks, workspace
finalization, native Grok support, and both sets of tests are retained.
- Current-head Greptile: **5/5**, with no blocking findings. The review
names this exact commit. All **59 reported checks are terminal: 55
successful, 4 skipped, zero pending or failing**. This includes the full
root general and serialized suites, separate runner checks, typecheck,
build, canary, browser E2E, Docker, and security checks. The successful
legacy security status is included in that total.
- After integration: workspace typecheck and full build passed. Separate
runner checks passed: **2,160 TypeScript tests (10 skipped), 582 Rust
tests, and 39 preparation checks**. Other passing checks include 621
Product E2E harness units, 376 focused shared/adapter tests, 160
real-database/API tests, 86 Hermes tests, 18 browser-support checks, and
Product E2E typechecking. The complete root suite passed in CI. The
duplicate local monolithic root run was stopped after that CI result; it
is not counted as a completed local pass.
- New native recovery coverage retains full assignment, completion
contract, and explicit skill selection after safe replacement, for old
and prepared input formats. Full native session test file: **136/136
passed**.
- New Claude ACP coverage captures actual fresh, resumed, and
missing-session fallback requests. It verifies one assignment copy,
comment order, identical text under distinct comment IDs, and full
fallback context. Full file: **33/33 passed**. Both affected TypeScript
checks passed.
- Existing deterministic tests cover source revisions, approval and
trust boundaries, completion validation, custom templates, compatible
sessions, standalone driver wrapping, and maintained adapter transport
requests.
- Provider-free browser support: **17/17 passed** after the master
merge. Service-worker unit tests: **33/33 passed**. The module-overload
regression failed before the repair and passed after it in real
Chromium.

### Fresh live comparisons

The new batch ran exactly four Product E2E attempts. **All four passed
on the first attempt; no retries.** Each has six terminal matchers plus
the existing browser lifecycle and invariant checks.

| Exact case ID | Control | Candidate |
|---|---|---|
| `core-compatibility.runner-codex.local.plan-revise-accept` | Passed |
Passed |
|
`local-session-integrity.runner-acpx-claude.local.structured-question-restart-resume`
| Passed | Passed |

The plan case checks a revised canonical plan and revision-bound
approval before completion. The question case restarts the server before
submitting the answer, then verifies the continuation completes.

Control source is `dfa4e1bda8d50a1a01746603251a9128dbe9d0d6`. Candidate
source is `79fcdb5dece501d28064ea9da306603881b46f0c`. They use identical
frozen definitions and provider versions: Codex `0.156.0` with
`gpt-5.6-sol`; ACPX `0.13.1` / Claude ACP `0.73.0` with
`claude-sonnet-5`. The September 24 head added master browser recovery
and test-only changes. The September 28 head also integrates newer
master changes, including cancellation, workspace finalization, and
native Grok. These are frozen-source live results, not exact-head live
runs.

The candidate received one description copy where the control initially
received three. The submitted initial plan envelopes were 7,969 versus
19,097 characters. Question envelopes were 7,592 versus 18,919. These
are structural measurements, not whole-provider token or dollar savings.

### Earlier evidence and failed attempts

- The preceding fresh batch has four effective passing pairs: OpenCode
comment continuation and assigned skill, native Codex comment
continuation, and native Claude comment continuation. It retains **11
attempts: eight passed and three failed**.
- Original failures remain recorded: missing local PostgreSQL library
links before task creation; host-sleep cleanup after task/page checks
passed; and a Claude **control** session-open rejection before a model
turn. Setup was repaired identically on both worktrees. The permitted
unchanged infrastructure retries passed. The underlying Claude provider
startup error was not retained and remains unknown.
- Older R2 retains **17 passes and one failure** across 18 attempts,
including eight both-pass native/legacy Codex/Claude pairs. Its OpenCode
blank-page failure led to the service-worker repair. R2 is historical
evidence: master changed the native fixed prompt and removed duplicate
wake environment data afterward.
- The September 24 CI run initially failed one unrelated preview
readiness test (`ECONNREFUSED` on its local fixture). Its test and
production code match master. Isolated local verification passed **28
tests, 3 skipped**. One unchanged CI retry passed the full shard: **831
passed, 1 skipped**, including all **31 preview-exposure tests**. The
aggregate CI gate passed afterward. The precise startup cause remains
unknown; a port race is a hypothesis, not a proved cause.

### Limits

The original wider profile/workflow matrix, repeated trials, and remote
Daytona qualification are incomplete. These results support a focused
merge recommendation, not statistical equivalence or universal harness
qualification. Some usage receipts are missing in both variants, so no
token or dollar savings are claimed. The $500 ceiling was preserved
using conservative allowances; failed attempts and unknown charges
remain in the ledger.

Reproduce the focused additions with `pnpm exec vitest run
packages/adapters/claude-local/src/server/acp.test.ts` and `pnpm
--filter @paperclipai/paperclip-runner exec vitest run
src/native-session-runtime.test.ts`. Full checks use `pnpm -r
typecheck`, `pnpm test:run`, `pnpm build`, and the separate runner
checks. Paid evals require the frozen definitions, profiles, and
credentials; do not use `--all` as a substitute for the selected cases.

## Risks

- Context placement changes can affect model behavior. Deterministic
checks cover the selected paths, but live qualification remains
incomplete.
- The stale-question guard can reject a request when queued human
comments arrived during the run. This is intended.
- New stored inputs and model envelopes retain compatibility readers for
older active sessions.
- Custom templates may intentionally repeat content.
- Removing a model-facing working-directory sentence does not change
filesystem, command, sandbox, or permission configuration.
- The worker bypass applies only to same-origin development module
paths. Cache-policy tests preserve private-response handling and
production asset caching. Mounted HTTP fixture changes remain test-only.
- This PR does not claim measured token savings or statistical
equivalence across every harness.

## Model Used

OpenAI Codex, exact model gpt-6-astra, with repository tools and code
execution. Bounded supporting work used gpt-5.6-luna and gpt-6-luna. The
serving context-window size is not exposed in this task.

## 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 described the issue in-PR using the required issue fields
- [x] I have not referenced internal/instance-local Paperclip issues or
links
- [x] My branch name describes the change and contains no internal
ticket id
- [x] I have run the focused local checks and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect these changes
- [x] I have considered and documented risks above
- [x] All current-head Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
for the current head
- [x] I will address all Greptile and reviewer comments before
requesting merge

---------

Co-authored-by: Paperclip <noreply@paperclip.ing>
2026-09-28 14:49:14 -05:00
..
…

Paperclip Adapters for Hermes Agent

A Paperclip adapter package that lets you run Hermes Agent as a managed employee in a Paperclip company.

Hermes Agent is a full-featured AI agent by Nous Research with 30+ native tools, persistent memory, session persistence, 80+ skills, MCP support, and multi-provider model access.

This package owns both built-in Hermes adapter types:

  • hermes_local runs the local Hermes CLI as a child process. The package root exports remain compatible with the original local adapter.
  • hermes_gateway calls an already-running Hermes API server over HTTP/SSE. Gateway entrypoints live under the ./gateway export namespace.

Choose hermes_local when Paperclip and Hermes run on the same trusted host and Paperclip should start hermes chat for each heartbeat. Choose hermes_gateway when Hermes is already running as an API server, often on another host, in Docker, or behind a private-network/TLS endpoint. The adapter type keys did not change during package consolidation.

Key Features

This adapter provides:

  • 8 inference providers — Anthropic, OpenRouter, OpenAI, Nous, OpenAI Codex, ZAI, Kimi Coding, MiniMax
  • Skills integration — Scans both Paperclip-managed and Hermes-native skills (~/.hermes/skills/), with sync/list/resolve APIs
  • Structured transcript parsing — Raw Hermes stdout is parsed into typed TranscriptEntry objects so Paperclip renders proper tool cards with status icons and expand/collapse
  • Rich post-processing — Converts Hermes ASCII banners, setext headings, and +--+ table borders into clean GFM markdown
  • Comment-driven wakes — Agents wake to respond to issue comments, not just task assignments
  • Auto model detection — Reads ~/.hermes/config.yaml to pre-populate the UI with the user's configured model
  • Session codec — Structured validation and migration of session state across heartbeats
  • Benign stderr reclassification — MCP init messages and structured logs are reclassified so they don't appear as errors in the UI
  • Session source tagging — Sessions are tagged as tool source so they don't clutter the user's interactive history
  • Filesystem checkpoints — Optional --checkpoints for rollback safety
  • Thinking effort control — Passes --reasoning-effort for thinking/reasoning models

Hermes Agent Capabilities

Feature Claude Code Codex Hermes Agent
Persistent memory ❌ ❌ ✅ Remembers across sessions
Native tools ~5 ~5 30+ (terminal, file, web, browser, vision, git, etc.)
Skills system ❌ ❌ ✅ 80+ loadable skills
Session search ❌ ❌ ✅ FTS5 search over past conversations
Sub-agent delegation ❌ ❌ ✅ Parallel sub-tasks
Context compression ❌ ❌ ✅ Auto-compresses long conversations
MCP client ❌ ❌ ✅ Connect to any MCP server
Multi-provider Anthropic only OpenAI only ✅ 8 providers out of the box

Installation

This package ships with Paperclip core as the built-in hermes_local and hermes_gateway adapters. No Adapter manager installation is required for normal Paperclip use.

Prerequisites

  • Hermes Agent installed (pip install hermes-agent)
  • Python 3.10+
  • At least one LLM API key (Anthropic, OpenRouter, or OpenAI)

Quick Start

1. Optional: override the built-in for adapter development

For local adapter development, install the package from a local path in Adapter manager, or add an entry to ~/.paperclip/adapter-plugins.json and restart Paperclip. The external package can override either built-in Hermes adapter while it is enabled:

[
  {
    "packageName": "@paperclipai/hermes-paperclip-adapter",
    "localPath": "/absolute/path/to/paperclip/packages/adapters/hermes",
    "type": "hermes_local",
    "installedAt": "2026-06-23T00:00:00.000Z"
  }
]

Use "type": "hermes_gateway" with the same package when testing a gateway override.

The package root exports createServerAdapter() for the local server adapter, a declarative config schema for the generic agent form, and ./ui-parser for local run transcript parsing. Gateway entrypoints are exported from ./gateway, ./gateway/server, ./gateway/ui, ./gateway/cli, and ./gateway/ui-parser. Paperclip core imports these same package entrypoints for built-in adapter registration.

2. Create a local Hermes agent in Paperclip

In the Paperclip UI or via API, create an agent with adapter type hermes_local:

{
  "name": "Hermes Engineer",
  "adapterType": "hermes_local",
  "adapterConfig": {
    "model": "anthropic/claude-sonnet-4",
    "maxIterations": 50,
    "timeoutSec": 300,
    "persistSession": true,
    "enabledToolsets": ["terminal", "file", "web"]
  }
}

This mode shells out to the local hermes CLI. Paperclip injects runtime environment variables and captures stdout/stderr from the child process.

3. Create a Hermes gateway agent in Paperclip

Start Hermes with its API server enabled first:

API_SERVER_ENABLED=true \
API_SERVER_KEY=<generated-secret> \
hermes gateway run --replace --accept-hooks

Then create an agent with adapter type hermes_gateway:

{
  "name": "Hermes Gateway Engineer",
  "adapterType": "hermes_gateway",
  "adapterConfig": {
    "apiBaseUrl": "http://127.0.0.1:8642",
    "apiKey": "<same-value-as-API_SERVER_KEY>",
    "paperclipApiUrl": "http://127.0.0.1:3100",
    "sessionKeyStrategy": "issue",
    "timeoutSec": 600
  }
}

If the URL you have is the default Hermes dashboard at http://127.0.0.1:9119 or the default chat URL at http://127.0.0.1:9119/chat, Paperclip maps it to http://127.0.0.1:9119/api before calling Hermes API routes. /chat and the dashboard root are browser UI routes; Paperclip tests /api/health and starts runs with /api/v1/runs after mapping them to the API base.

This mode does not start Hermes. It creates runs with POST /v1/runs, streams Hermes events with SSE, polls run status as a fallback, and stops timed-out runs with POST /v1/runs/{run_id}/stop.

Compatibility with the old gateway package

@paperclipai/adapter-hermes-gateway remains as a deprecated compatibility shim for one release. It re-exports the gateway entrypoints from @paperclipai/hermes-paperclip-adapter/gateway and preserves the legacy exports for existing plugin installs. New installs and built-in Paperclip registrations should use @paperclipai/hermes-paperclip-adapter; the adapter type remains hermes_gateway.

Runtime API guidance

Hermes receives Paperclip runtime identity through environment variables:

  • PAPERCLIP_API_URL
  • PAPERCLIP_API_KEY
  • PAPERCLIP_RUN_ID

Prompts should reference those variables directly. Command output may redact secret values, so do not copy printed tokens into comments or config. Use Authorization: Bearer $PAPERCLIP_API_KEY on Paperclip API requests and X-Paperclip-Run-Id: $PAPERCLIP_RUN_ID on mutating issue requests. For multiline comments or status updates, preserve newlines with a heredoc plus jq --arg.

Hermes-originated Paperclip tasks

The package includes a Hermes skill/helper for the reverse direction: a user starts in Hermes and asks Hermes to create or update Paperclip work. This is not the same as Paperclip waking Hermes through hermes_local or hermes_gateway.

Configure Paperclip access in Hermes env/profile secrets, not prompt text:

PAPERCLIP_API_URL=http://127.0.0.1:3100/api
PAPERCLIP_BRIDGE_API_KEY=<task-bridge-scoped-agent-api-key>

Optional env values:

  • PAPERCLIP_COMPANY_ID
  • PAPERCLIP_AGENT_ID
  • PAPERCLIP_RUN_ID

The bundled paperclip-task-bridge skill provides deterministic helper commands:

node ./paperclip-task.mjs list-assigned
node ./paperclip-task.mjs create-task --parent-id "<approved-parent-issue-id>" --title "Investigate checkout failures" --description "Capture failing request and root cause."
node ./paperclip-task.mjs comment --issue PAP-123 --body "Found the failing request path."
node ./paperclip-task.mjs update-status --issue PAP-123 --status in_review --comment "Ready for review."

The helper reads credentials from environment variables and prints only JSON summaries. It supports create-task, comment, update-status, and list-assigned.

Create the bridge key with scope.kind = "task_bridge" plus a parentIssueId or projectId boundary. Do not use a normal claimed agent API key for internet-facing Hermes chat/webhook task-bridge operations.

4. Assign work

Create issues in Paperclip and assign them to your Hermes agent. On each heartbeat, Hermes will:

  1. Receive the task instructions
  2. Use its full tool suite to complete the work
  3. Report results back to Paperclip
  4. Persist session state for continuity

Configuration Reference

Core

Field Type Default Description
model string anthropic/claude-sonnet-4 Model in provider/model format
provider string (auto-detected) API provider: auto, openrouter, nous, openai-codex, zai, kimi-coding, minimax, minimax-cn
timeoutSec number 300 Execution timeout in seconds
graceSec number 10 Grace period before SIGKILL

Tools

Field Type Default Description
toolsets string (all) Comma-separated toolsets to enable (e.g. "terminal,file,web")

Available toolsets: terminal, file, web, browser, code_execution, vision, mcp, creative, productivity

Session & Workspace

Field Type Default Description
persistSession boolean true Resume sessions across heartbeats
worktreeMode boolean false Git worktree isolation
checkpoints boolean false Enable filesystem checkpoints for rollback

Advanced

Field Type Default Description
hermesCommand string hermes Custom CLI binary path
verbose boolean false Enable verbose output
quiet boolean true Quiet mode (clean output, no banner/spinner)
extraArgs string[] [] Additional CLI arguments
env object {} Extra environment variables
promptTemplate string (built-in) Custom prompt template
paperclipApiUrl string http://127.0.0.1:3100/api Paperclip API base URL

Prompt Template Variables

Use {{variable}} syntax in promptTemplate:

Variable Description
{{agentId}} Paperclip agent ID
{{agentName}} Agent display name
{{companyId}} Company ID
{{companyName}} Company name
{{runId}} Current heartbeat run ID
{{taskId}} Assigned task/issue ID
{{taskTitle}} Task title
{{taskBody}} Task instructions
{{projectName}} Project name
{{paperclipApiUrl}} Paperclip API base URL
{{commentId}} Comment ID (when woken by a comment)
{{wakeReason}} Reason this run was triggered

Conditional sections:

  • {{#taskId}}...{{/taskId}} — included only when a task is assigned
  • {{#noTask}}...{{/noTask}} — included only when no task (heartbeat check)
  • {{#commentId}}...{{/commentId}} — included only when woken by a comment

Architecture

Paperclip                          Hermes Agent
┌──────────────────┐               ┌──────────────────┐
│  Heartbeat       │               │                  │
│  Scheduler       │───execute()──▶│  hermes chat -q  │
│                  │               │                  │
│  Issue System    │               │  30+ Tools       │
│  Comment Wakes   │◀──results─────│  Memory System   │
│                  │               │  Session DB      │
│  Cost Tracking   │               │  Skills          │
│                  │               │  MCP Client      │
│  Skill Sync      │◀──snapshot────│  ~/.hermes/skills│
│  Org Chart       │               │                  │
└──────────────────┘               └──────────────────┘

The adapter spawns Hermes Agent's CLI in single-query mode (-q). Hermes processes the task using its full tool suite, then exits. The adapter:

  1. Captures stdout/stderr and parses token usage, session IDs, and cost
  2. Parses raw output into structured TranscriptEntry objects (tool cards with status icons)
  3. Post-processes Hermes ASCII formatting (banners, setext headings, table borders) into clean GFM markdown
  4. Reclassifies benign stderr (MCP init, structured logs) so they don't show as errors
  5. Tags sessions as tool source to keep them separate from interactive usage
  6. Reports results back to Paperclip with cost, usage, and session state

Session persistence works via Hermes's --resume flag — each run picks up where the last one left off, maintaining conversation context, memories, and tool state across heartbeats. The sessionCodec validates and migrates session state between runs.

Skills Integration

The adapter scans two skill sources and merges them:

  • Paperclip-managed skills — bundled with the adapter, togglable from the UI
  • Hermes-native skills — from ~/.hermes/skills/, read-only, always loaded

The listSkills / syncSkills APIs expose a unified snapshot so the Paperclip UI can display both managed and native skills in one view.

Development

git clone https://github.com/paperclipai/paperclip
cd paperclip/packages/adapters/hermes
pnpm install
pnpm build

License

MIT — see LICENSE