public/chiche-plugins
Watch
1
0
Fork
You've already forked chiche-plugins
 0
Code Issues Pull requests Activity Actions

refactor: delegate fj work to the fj-worker subagent instead of forking #8

Merged
jerome merged 4 commits from refactor/forgejo-cli-delegation into main 2026-06-25 09:08:09 +03:00
Conversation 3 Commits 4 Files changed 4 +595 -286
claude commented 2026-06-25 08:47:16 +03:00
Collaborator
Copy link

Goal

Invert the forgejo-cli plugin from an auto-forking skill into a thin router
that splits Forgejo work two ways:

  • Reads (PR/issue body, comments, formal reviews, inline code comments) run
    a deterministic bundled script (scripts/fj_read.py) -- no model in the loop,
    so content comes back exactly as written (a body that is a spec stays verbatim).
  • Writes + the long tail delegate to the fj-worker subagent (Sonnet),
    which holds the full fj reference + footguns and runs one small,
    self-contained task at a time, taking any multi-line content via an absolute
    tmp-file path the caller writes.

Either way the main context stays lean and carries zero fj mechanics.

Why

The previous design used context: fork so any fj mention forked the whole
task into the worker. Two problems: the main context did no thinking of its own,
and routing a content read through an LLM risked summarising/truncating a body
that is actually a spec. The new model keeps the main context in control, makes
reads deterministic (a script, not a model), and reserves the worker for the
footgun-heavy writes and adaptive error recovery.

Steps

  • Move the canonical fj reference (commands, footguns, REST fallbacks,
    token scopes, subcommand surface) into the fj-worker agent, and add an
    operating contract: one small self-contained task per dispatch; large
    content arrives as an absolute tmp-file path to Read; results terse,
    but returned content (a body) verbatim.
  • Move the extras.md companion next to the agent
    (agents/references/extras.md) and retarget its cross-references at the
    agent reference, since the skill no longer owns it.
  • Shrink SKILL.md to a thin router: drop the context: fork + agent:
    frontmatter, trim the description to a pure trigger, and make the body the
    two-path contract (script the reads, delegate the writes; the tmp-file rule
    for bodies; what comes back).
  • Add scripts/fj_read.py -- a stdlib-only script that reads fj's stored
    token and hits the Forgejo REST API to dump a PR/issue verbatim (metadata,
    body, flat comments, and reviews + inline comments, which fj can't show)
    to a tmp file or stdout (--format json for machine output). Route reads
    through it in the skill router.

Notes

  • Reference lives only in the agent (agents/fj-worker.md +
    agents/references/extras.md); main context carries just the small router body
    when the skill triggers.
  • Reads are deterministic and free of a subagent round-trip: fj_read.py is more
    capable than fj for reads (reviews/inline comments are REST-only; it also
    sidesteps the broken fj pr status).
  • No automated test suite for this plugin repo. fj_read.py was verified live
    against this PR (#8): body verbatim, comments captured, JSON + --out + error
    paths (bad host, 404) all behave.
  • Follow-up (separate PR, out of scope): the taskpy suite still tells callers
    to "invoke the forgejo-cli skill ... for syntax" and references "the
    forgejo-cli skill's review subsection" -- those now resolve to the agent /
    fj_read.py, so the references want updating.

Depends on: none.

## Goal Invert the `forgejo-cli` plugin from an auto-forking skill into a **thin router** that splits Forgejo work two ways: - **Reads** (PR/issue body, comments, formal reviews, inline code comments) run a deterministic bundled script (`scripts/fj_read.py`) -- no model in the loop, so content comes back exactly as written (a body that is a spec stays verbatim). - **Writes + the long tail** delegate to the `fj-worker` subagent (Sonnet), which holds the full `fj` reference + footguns and runs one small, self-contained task at a time, taking any multi-line content via an absolute tmp-file path the caller writes. Either way the main context stays lean and carries zero `fj` mechanics. ## Why The previous design used `context: fork` so any `fj` mention forked the *whole* task into the worker. Two problems: the main context did no thinking of its own, and routing a content read through an LLM risked summarising/truncating a body that is actually a spec. The new model keeps the main context in control, makes reads deterministic (a script, not a model), and reserves the worker for the footgun-heavy writes and adaptive error recovery. ## Steps - [x] Move the canonical `fj` reference (commands, footguns, REST fallbacks, token scopes, subcommand surface) into the `fj-worker` agent, and add an operating contract: one small self-contained task per dispatch; large content arrives as an absolute tmp-file path to `Read`; results terse, but returned *content* (a body) verbatim. - [x] Move the `extras.md` companion next to the agent (`agents/references/extras.md`) and retarget its cross-references at the agent reference, since the skill no longer owns it. - [x] Shrink `SKILL.md` to a thin router: drop the `context: fork` + `agent:` frontmatter, trim the description to a pure trigger, and make the body the two-path contract (script the reads, delegate the writes; the tmp-file rule for bodies; what comes back). - [x] Add `scripts/fj_read.py` -- a stdlib-only script that reads `fj`'s stored token and hits the Forgejo REST API to dump a PR/issue verbatim (metadata, body, flat comments, and reviews + inline comments, which `fj` can't show) to a tmp file or stdout (`--format json` for machine output). Route reads through it in the skill router. ## Notes - Reference lives **only** in the agent (`agents/fj-worker.md` + `agents/references/extras.md`); main context carries just the small router body when the skill triggers. - Reads are deterministic and free of a subagent round-trip: `fj_read.py` is more capable than `fj` for reads (reviews/inline comments are REST-only; it also sidesteps the broken `fj pr status`). - No automated test suite for this plugin repo. `fj_read.py` was verified live against this PR (#8): body verbatim, comments captured, JSON + `--out` + error paths (bad host, 404) all behave. - Follow-up (separate PR, out of scope): the `taskpy` suite still tells callers to "invoke the `forgejo-cli` skill ... for syntax" and references "the forgejo-cli skill's review subsection" -- those now resolve to the agent / `fj_read.py`, so the references want updating. **Depends on:** none.
The worker now executes one small, self-contained fj operation per
dispatch from its own reference, reading large content (a PR/issue body)
from an absolute tmp-file path. Move the extras companion next to the
agent (agents/references/extras.md) since the skill no longer owns it.
refactor: make forgejo-cli a thin router that delegates to fj-worker
All checks were successful
gitleaks / gitleaks (pull_request) Successful in 3s
Details
92849783ef 
Drop context:fork and the auto-fork frontmatter; the skill body now
fits in the main context and instructs it to delegate each fj operation
to the fj-worker subagent as a small self-contained task, passing any
multi-line body through a tmp file. Main context no longer carries the
fj reference.
claude commented 2026-06-25 08:47:49 +03:00
Author
Collaborator
Copy link

2026-06-25 -- resolver pass on refactor/forgejo-cli-delegation at 9284978.

  • Steps done: all 3 (reference + operating contract into the agent; extras moved to agents/references/; SKILL.md shrunk to the delegation router, context:fork dropped).
  • Deferred: none (no automated test suite for this plugin repo; verification is manual -- delegate a real fj op through the worker and confirm the tmp-file body path works end to end).
  • The fj reference now lives only in agents/fj-worker.md + agents/references/extras.md; main context carries just the small router body when the skill triggers.
2026-06-25 -- resolver pass on refactor/forgejo-cli-delegation at 9284978. - Steps done: all 3 (reference + operating contract into the agent; extras moved to agents/references/; SKILL.md shrunk to the delegation router, context:fork dropped). - Deferred: none (no automated test suite for this plugin repo; verification is manual -- delegate a real fj op through the worker and confirm the tmp-file body path works end to end). - The fj reference now lives only in agents/fj-worker.md + agents/references/extras.md; main context carries just the small router body when the skill triggers.
Route reads (body, flat comments, formal reviews, inline comments)
through a stdlib-only script that hits the Forgejo REST API with fj's
stored token and emits the record verbatim to a tmp file -- no model in
the loop, so a body that is a spec comes back exactly as written, and
reviews/inline comments (which fj can't show) are included. The skill
router now scripts reads and delegates only writes + the long tail to
fj-worker; the worker keeps a verbatim-content safeguard for the
fallback case.
Merge remote-tracking branch 'origin/main' into refactor/forgejo-cli-delegation
All checks were successful
gitleaks / gitleaks (pull_request) Successful in 3s
Details
1264e3f98c
claude commented 2026-06-25 09:03:45 +03:00
Author
Collaborator
Copy link

2026-06-25 -- second resolver pass on refactor/forgejo-cli-delegation at 1264e3f.

  • Added scripts/fj_read.py (Step 4): deterministic, stdlib-only REST dump of a PR/issue -- body + flat comments + reviews + inline comments, verbatim, to a tmp file or stdout (--format json). Reads no longer route through fj-worker.
  • SKILL.md router now splits the two paths (script the reads, delegate the writes); fj-worker keeps a verbatim-content safeguard for the fallback read.
  • Verified fj_read.py live against this PR: body verbatim, comments captured, JSON/--out/error paths (bad host, 404) all behave.
  • Caught the branch up to origin/main (PR #7's taskpy rename landed while this was in flight); clean merge, diff stays scoped to the 4 forgejo-cli files.
  • Follow-up (separate PR): taskpy-implement still points callers at "the forgejo-cli skill ... for syntax" / "review subsection" (SKILL.md:165,:315; worktree-and-branches.md:53,:280) -- those now resolve to the agent + fj_read.py.
2026-06-25 -- second resolver pass on refactor/forgejo-cli-delegation at 1264e3f. - Added scripts/fj_read.py (Step 4): deterministic, stdlib-only REST dump of a PR/issue -- body + flat comments + reviews + inline comments, verbatim, to a tmp file or stdout (--format json). Reads no longer route through fj-worker. - SKILL.md router now splits the two paths (script the reads, delegate the writes); fj-worker keeps a verbatim-content safeguard for the fallback read. - Verified fj_read.py live against this PR: body verbatim, comments captured, JSON/--out/error paths (bad host, 404) all behave. - Caught the branch up to origin/main (PR #7's taskpy rename landed while this was in flight); clean merge, diff stays scoped to the 4 forgejo-cli files. - Follow-up (separate PR): taskpy-implement still points callers at "the forgejo-cli skill ... for syntax" / "review subsection" (SKILL.md:165,:315; worktree-and-branches.md:53,:280) -- those now resolve to the agent + fj_read.py.
jerome approved these changes 2026-06-25 09:08:02 +03:00
jerome changed title from WIP: refactor: delegate fj work to the fj-worker subagent instead of forking to refactor: delegate fj work to the fj-worker subagent instead of forking 2026-06-25 09:08:04 +03:00
jerome merged commit c7093d7c35 into main 2026-06-25 09:08:09 +03:00
Sign in to join this conversation.
Reviewers
No reviewers
jerome
Labels
Clear labels   
bug
Something is not working

  
duplicate
This issue or pull request already exists

  
enhancement
New feature

  
help wanted
Need some help

  
invalid
Something is wrong

  
question
More information is needed

  
wontfix
This won't be fixed

No labels
bug
duplicate
enhancement
help wanted
invalid
question
wontfix
Milestone
Clear milestone
No items
No milestone
Assignees
Clear assignees
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
public/chiche-plugins!8
No description provided.