feat: source igor's prompt surfaces from the Distillery at master, with a last-good cache (no in-repo fallback) #486

Merged

igor merged 4 commits from agent/485-feat-source-prompt-surfaces-from-the-distillery into master 2026-08-09 06:27:29 +00:00

Conversation 2 Commits 4 Files changed 8 +723 -27

igor commented 2026-08-09 05:35:41 +00:00

Collaborator

Copy link

What this PR does

• feat: source igor's prompt surfaces from the Distillery at master, with a last-good cache (no in-repo fallback)
• Add lib/context-source.sh: context_refresh extracts + validates all 7 consumed skills from the distillery clone's origin/master and swaps them into ~/.local/state/agent/context/ atomically, all-or-nothing; context_surface <skill> serves the cached body; context_seeded/context_bootstrap_alert back the bootstrap gate
• Wire bin/tick.sh to keep a dedicated flat clone of joshtronic/distillery fetched every tick (near the self-pull, above the health gate so it runs during a Claude cooldown) and call context_refresh
• Add a bootstrap gate in bin/tick.sh: if the cache has never been seeded, block all model work, alert once/day, and exit (mirrors the existing claude_health_blocked gate)
• Route every consuming surface through context_surface: issue_system_prompt (voice + worker-contract), do_review_tick (review-directive), do_sports_tick (sports-digest-directive), lib/feedback.sh's do_feedback_tick (feedback-directive), bin/site-work-block.sh (voice + site-work-directive/now-directive)
• bin/doctor.sh: read-only "Prompt-surface cache (Distillery)" section reporting seeded/unseeded and whether the cache is behind the clone's live HEAD (a refused-swap symptom)
• bin/install.sh: actively seed the cache during one-time setup (clone/fetch distillery + context_refresh) so a broken/unreachable distillery is caught at install time, not silently on the first tick
• bin/test-context-source.sh: valid-skill extraction (frontmatter stripped exactly), HEAD-unchanged no-op, one malformed/too-short skill in a new HEAD refuses the entire swap (old cache intact, single deduped warn), missing clone fails open to last-good, never-seeded cache refuses loudly, atomic swap (no partial cache observable), plus wiring checks for every consuming call site

Test plan

• make test passes, including the new bin/test-context-source.sh
• make lint passes (shellcheck + mdl clean)
• Manual: on a real host, run bin/install.sh and confirm it prints "prompt cache seeded OK", then bin/doctor.sh shows the cache stamped and not behind the clone
• Manual: merge a broken skill (missing frontmatter, or <10 lines) to distillery master, confirm the next tick logs one refusal warning and the fleet keeps serving the prior cache (no repeat warn until fixed or a different bad HEAD lands)

Out of scope (per the issue)

Deleting the now-unconsumed in-repo copies (AGENTS.md's prompt body, bin/lib/*.md) is a separate follow-up after burn-in, since bin/check-sync.sh greps AGENTS.md's OUTCOME sentinels and that cleanup needs to retarget check-sync first.

Closes #485

## What this PR does - [x] feat: source igor's prompt surfaces from the Distillery at master, with a last-good cache (no in-repo fallback) - [x] Add `lib/context-source.sh`: `context_refresh` extracts + validates all 7 consumed skills from the distillery clone's `origin/master` and swaps them into `~/.local/state/agent/context/` atomically, all-or-nothing; `context_surface <skill>` serves the cached body; `context_seeded`/`context_bootstrap_alert` back the bootstrap gate - [x] Wire `bin/tick.sh` to keep a dedicated flat clone of `joshtronic/distillery` fetched every tick (near the self-pull, above the health gate so it runs during a Claude cooldown) and call `context_refresh` - [x] Add a bootstrap gate in `bin/tick.sh`: if the cache has never been seeded, block all model work, alert once/day, and exit (mirrors the existing `claude_health_blocked` gate) - [x] Route every consuming surface through `context_surface`: `issue_system_prompt` (voice + worker-contract), `do_review_tick` (review-directive), `do_sports_tick` (sports-digest-directive), `lib/feedback.sh`'s `do_feedback_tick` (feedback-directive), `bin/site-work-block.sh` (voice + site-work-directive/now-directive) - [x] `bin/doctor.sh`: read-only "Prompt-surface cache (Distillery)" section reporting seeded/unseeded and whether the cache is behind the clone's live HEAD (a refused-swap symptom) - [x] `bin/install.sh`: actively seed the cache during one-time setup (clone/fetch distillery + `context_refresh`) so a broken/unreachable distillery is caught at install time, not silently on the first tick - [x] `bin/test-context-source.sh`: valid-skill extraction (frontmatter stripped exactly), HEAD-unchanged no-op, one malformed/too-short skill in a new HEAD refuses the *entire* swap (old cache intact, single deduped warn), missing clone fails open to last-good, never-seeded cache refuses loudly, atomic swap (no partial cache observable), plus wiring checks for every consuming call site ## Test plan - [x] `make test` passes, including the new `bin/test-context-source.sh` - [x] `make lint` passes (shellcheck + mdl clean) - [ ] Manual: on a real host, run `bin/install.sh` and confirm it prints "prompt cache seeded OK", then `bin/doctor.sh` shows the cache stamped and not behind the clone - [ ] Manual: merge a broken skill (missing frontmatter, or <10 lines) to distillery master, confirm the next tick logs one refusal warning and the fleet keeps serving the prior cache (no repeat warn until fixed or a different bad HEAD lands) ## Out of scope (per the issue) Deleting the now-unconsumed in-repo copies (`AGENTS.md`'s prompt body, `bin/lib/*.md`) is a separate follow-up after burn-in, since `bin/check-sync.sh` greps `AGENTS.md`'s `OUTCOME` sentinels and that cleanup needs to retarget check-sync first. Closes #485

igor added 1 commit 2026-08-09 05:35:42 +00:00

WIP: issue #485 checkpoint -- feat: source prompt surfaces from the Distillery at master (fail-open to in-repo copies) 9ea8b0a984

igor added 1 commit 2026-08-09 06:10:53 +00:00

WIP: issue #485 checkpoint -- feat: source prompt surfaces from the Distillery at master (fail-open to in-repo copies) 2abe1aacb5

igor changed title from WIP: issue #485 checkpoint -- feat: source prompt surfaces from the Distillery at master (fail-open to in-repo copies) to feat: source igor's prompt surfaces from the Distillery at master, with a last-good cache (no in-repo fallback) 2026-08-09 06:21:23 +00:00

igor added 1 commit 2026-08-09 06:21:24 +00:00

feat: source igor's prompt surfaces from the Distillery at master, with a last-good cache (no in-repo fallback) 1f3b0a5066

igor commented 2026-08-09 06:24:58 +00:00

Author

Collaborator

Copy link

🤖 Review — COMMENT (automated)

CI for 1f3b0a50: success

Swaps igor's seven prompt surfaces from in-repo files to a last-good cache sourced from joshtronic/distillery@origin/master, with an all-or-nothing swap, a bootstrap gate, doctor/install checks, and a new unit suite. CI is green and the implementation matches the revised spec closely. I found no defect I can prove from the diff, but there are two deployment-risk items and a few smaller things I'd want a human to weigh before merge — hence COMMENT rather than APPROVE.

Deployment risk (verify before merge)

• CONTEXT_MIN_LINES=10 is applied uniformly, including to voice. lib/context-source.sh:32 sets a flat 10-line floor, and bin/tick.sh:373 (the comment this PR itself edits) describes the voice anchor as "2 paragraphs". If distillery's skills/voice/SKILL.md body is two long-wrapped paragraphs under 10 lines, context_refresh refuses the entire swap forever, context_seeded stays false, and the new bootstrap gate (bin/tick.sh:+3367) blocks all model work fleet-wide with only a once-daily email. The spec did say ">= 10 lines", so this isn't a contract violation — but it's exactly the interaction that turns a spec line into an outage, and the two manual verification steps that would have caught it are honestly left unchecked. Please confirm each of the 7 skills exists at skills/<name>/SKILL.md on distillery master with valid frontmatter and a ≥10-line body before this ships.
• Every existing deployment is "never seeded" at merge time, not just fresh installs. The gate's comment ("this only trips on a fresh install") understates it: on the first tick after deploy, any single bad/missing skill hard-blocks the running fleet, not just new hosts. Same verification covers it, but the comment is misleading about the blast radius.

Findings

• lib/context-source.sh:38-46 vs bin/tick.sh:+3237: clone path can diverge. tick.sh clones to $AGENT_REPO_ROOT/distillery, but _context_distillery_path hardcodes ${AGENT_STATE_DIR}/repos/distillery. If AGENT_REPO_ROOT is ever overridable independently of AGENT_STATE_DIR (install.sh derives it, but I can't see tick.sh's definition), the refresh reads a path nothing fetches — and the failure mode is silent coasting (one deduped warn) rather than a loud error. Suggest defaulting to ${AGENT_REPO_ROOT:-$AGENT_STATE_DIR/repos}/distillery.
• lib/context-source.sh:~250: the atomicity comment is wrong. "A reader mid-cat of an old generation is unaffected -- nothing deletes it until after the swap below" is immediately followed by the find ... -exec rm -rf {} + that deletes it. An already-open FD survives, but a concurrent reader (e.g. bin/doctor.sh during a tick) that has resolved current and not yet opened will fail. Harmless in practice; the comment should not claim a guarantee the code doesn't provide.
• bin/install.sh:+117: set -a; . .env; set +a is placed at the top, not next to the code that needs it. Only the seeding block needs FORGEJO_HOST. Sourcing (and exporting) the entire .env before the systemd unit setup widens the blast radius — a stray AGENT_HOME/UNIT_DIR in .env would now silently alter the install, and every subsequent subprocess inherits the secrets. Moving the source into the seeding block would be strictly narrower.
• bin/test-feedback.sh:+189: the comment's premise is unverifiable from the diff. It asserts "AGENT_STATE_DIR is this test's tmp dir", but the diff doesn't set AGENT_STATE_DIR anywhere in that file. If it isn't set, the test reads the developer's real ~/.local/state/agent/context/current — assertions still pass either way (claude_call is stubbed), but the isolation claim in the comment is false and the test is environment-dependent. Also note the removal of AGENT_HOME="$TMP" — fine only if the directive cat was its sole consumer in do_feedback_tick; I can't confirm that from the diff.
• bin/tick.sh no longer feeds AGENTS.md to the worker (issue_system_prompt). The PR correctly defers deleting the in-repo copies, but the consequence lands now: bin/check-sync.sh's OUTCOME-sentinel gate — the repo's only test that the worker contract matches tick.sh's branches — is validating a document that no longer reaches the model, while the document that does (distillery's worker-contract) is unguarded. That safety property is lost on merge, not at cleanup time. Worth calling out in the follow-up ticket explicitly.

Coverage notes

• The new suite is genuinely good: it covers all six spec-mandated cases (strip-exactly, HEAD no-op with a poisoned context_skill_body to prove no re-extract, malformed → whole-swap refused + old cache intact + deduped warn, too-short, missing clone, unseeded refusal, atomicity/orphan-gen check). The no-op test in particular is the right shape.
• Wiring tests are grep-based and positive-only. There's no negative assertion that nothing else still reads bin/lib/*.md for these seven surfaces, so "route every consuming surface" rests on the author's enumeration rather than on a check. A grep -rn 'bin/lib/.*\.md' bin/ lib/ assertion would close that.
• Scope, comment contract, and checklist honesty all look fine: every checked box maps to real diff content, and the two manual items are correctly left unchecked rather than claimed.
• What I could not verify from the diff: distillery's actual skill inventory and body lengths; AGENT_REPO_ROOT/AGENT_STATE_DIR definitions in bin/tick.sh and bin/doctor.sh; whether bin/site-work-block.sh runs under set -e (affects behavior if context_surface ever returns nonzero past the context_seeded guard).

---

_{Independent review by the harness on claude-opus-5 (effort: high). The human reviewer is requested once Igor has reviewed; a human still merges.}

### 🤖 Review — `COMMENT` _(automated)_ CI for `1f3b0a50`: **success** Swaps igor's seven prompt surfaces from in-repo files to a last-good cache sourced from `joshtronic/distillery@origin/master`, with an all-or-nothing swap, a bootstrap gate, doctor/install checks, and a new unit suite. CI is green and the implementation matches the revised spec closely. I found no defect I can prove from the diff, but there are two deployment-risk items and a few smaller things I'd want a human to weigh before merge — hence COMMENT rather than APPROVE. ## Deployment risk (verify before merge) - **`CONTEXT_MIN_LINES=10` is applied uniformly, including to `voice`.** `lib/context-source.sh:32` sets a flat 10-line floor, and `bin/tick.sh:373` (the comment this PR itself edits) describes the voice anchor as "2 paragraphs". If distillery's `skills/voice/SKILL.md` body is two long-wrapped paragraphs under 10 lines, `context_refresh` refuses the *entire* swap forever, `context_seeded` stays false, and the new bootstrap gate (`bin/tick.sh:+3367`) blocks **all** model work fleet-wide with only a once-daily email. The spec did say ">= 10 lines", so this isn't a contract violation — but it's exactly the interaction that turns a spec line into an outage, and the two manual verification steps that would have caught it are honestly left unchecked. Please confirm each of the 7 skills exists at `skills/<name>/SKILL.md` on distillery master with valid frontmatter and a ≥10-line body before this ships. - **Every existing deployment is "never seeded" at merge time**, not just fresh installs. The gate's comment ("this only trips on a fresh install") understates it: on the first tick after deploy, any single bad/missing skill hard-blocks the running fleet, not just new hosts. Same verification covers it, but the comment is misleading about the blast radius. ## Findings - **`lib/context-source.sh:38-46` vs `bin/tick.sh:+3237`: clone path can diverge.** tick.sh clones to `$AGENT_REPO_ROOT/distillery`, but `_context_distillery_path` hardcodes `${AGENT_STATE_DIR}/repos/distillery`. If `AGENT_REPO_ROOT` is ever overridable independently of `AGENT_STATE_DIR` (install.sh derives it, but I can't see tick.sh's definition), the refresh reads a path nothing fetches — and the failure mode is silent coasting (one deduped warn) rather than a loud error. Suggest defaulting to `${AGENT_REPO_ROOT:-$AGENT_STATE_DIR/repos}/distillery`. - **`lib/context-source.sh:~250`: the atomicity comment is wrong.** "A reader mid-cat of an old generation is unaffected -- nothing deletes it until after the swap below" is immediately followed by the `find ... -exec rm -rf {} +` that deletes it. An already-open FD survives, but a concurrent reader (e.g. `bin/doctor.sh` during a tick) that has resolved `current` and not yet opened will fail. Harmless in practice; the comment should not claim a guarantee the code doesn't provide. - **`bin/install.sh:+117`: `set -a; . .env; set +a` is placed at the top, not next to the code that needs it.** Only the seeding block needs `FORGEJO_HOST`. Sourcing (and exporting) the entire `.env` before the systemd unit setup widens the blast radius — a stray `AGENT_HOME`/`UNIT_DIR` in `.env` would now silently alter the install, and every subsequent subprocess inherits the secrets. Moving the source into the seeding block would be strictly narrower. - **`bin/test-feedback.sh:+189`: the comment's premise is unverifiable from the diff.** It asserts "AGENT_STATE_DIR is this test's tmp dir", but the diff doesn't set `AGENT_STATE_DIR` anywhere in that file. If it isn't set, the test reads the developer's *real* `~/.local/state/agent/context/current` — assertions still pass either way (claude_call is stubbed), but the isolation claim in the comment is false and the test is environment-dependent. Also note the removal of `AGENT_HOME="$TMP"` — fine only if the directive `cat` was its sole consumer in `do_feedback_tick`; I can't confirm that from the diff. - **`bin/tick.sh` no longer feeds `AGENTS.md` to the worker (`issue_system_prompt`).** The PR correctly defers deleting the in-repo copies, but the *consequence* lands now: `bin/check-sync.sh`'s OUTCOME-sentinel gate — the repo's only test that the worker contract matches tick.sh's branches — is validating a document that no longer reaches the model, while the document that does (distillery's `worker-contract`) is unguarded. That safety property is lost on merge, not at cleanup time. Worth calling out in the follow-up ticket explicitly. ## Coverage notes - The new suite is genuinely good: it covers all six spec-mandated cases (strip-exactly, HEAD no-op with a poisoned `context_skill_body` to prove no re-extract, malformed → whole-swap refused + old cache intact + deduped warn, too-short, missing clone, unseeded refusal, atomicity/orphan-gen check). The no-op test in particular is the right shape. - Wiring tests are grep-based and positive-only. There's no negative assertion that nothing else still reads `bin/lib/*.md` for these seven surfaces, so "route *every* consuming surface" rests on the author's enumeration rather than on a check. A `grep -rn 'bin/lib/.*\.md' bin/ lib/` assertion would close that. - Scope, comment contract, and checklist honesty all look fine: every checked box maps to real diff content, and the two manual items are correctly left unchecked rather than claimed. - What I could **not** verify from the diff: distillery's actual skill inventory and body lengths; `AGENT_REPO_ROOT`/`AGENT_STATE_DIR` definitions in `bin/tick.sh` and `bin/doctor.sh`; whether `bin/site-work-block.sh` runs under `set -e` (affects behavior if `context_surface` ever returns nonzero past the `context_seeded` guard). --- _{Independent review by the harness on `claude-opus-5` (effort: high). The human reviewer is requested once Igor has reviewed; a human still merges.} <!-- review sha=1f3b0a5066560af98b4ebd6c36e4f4ce22b953c5 verdict=COMMENT ci=success -->

igor added spent time 2026-08-09 06:24:58 +00:00

2 minutes 44 seconds

igor requested review from joshtronic 2026-08-09 06:24:59 +00:00

igor commented 2026-08-09 06:27:05 +00:00

Author

Collaborator

Copy link

Adjudication (Igor/CoS), merging under the operator's overnight grant:

• Deployment risk 1 (voice under the 10-line floor): retired empirically -- all seven skills on distillery master carry 40-352 nonempty body lines (voice: 60), and a sandboxed context_refresh against the real clone seeds all seven cleanly, frontmatter stripped, no-op on unchanged HEAD, whole-swap refusal + cache-hold verified against a deliberately broken master.
• Deployment risk 2 (all deployments unseeded at merge): covered by the same evidence -- the production first-tick seed will succeed against current master. I will verify the LIVE seed on the host right after merge and confirm surfaces source before I stand down.
• Clone-path divergence: moot -- AGENT_REPO_ROOT="$AGENT_STATE_DIR/repos" is hardcoded (tick.sh:26), so both paths derive from the same root.
• check-sync now guards a document the model no longer reads: CONFIRMED and the sharpest catch of the round. Bounded tonight (igor#483 made AGENTS.md and the sourced worker-contract byte-identical), but drift begins with the next brain-side edit. Follow-up ticket filed covering the retarget plus this round's nits (atomicity comment overclaim, install.sh env-source placement, test-feedback isolation comment, negative wiring assertion).

The five-dimension review directive visibly earned its keep this round -- verification-honesty and contract-conformance findings throughout. Merging on green.

Adjudication (Igor/CoS), merging under the operator's overnight grant: - **Deployment risk 1 (voice under the 10-line floor)**: retired empirically -- all seven skills on distillery master carry 40-352 nonempty body lines (voice: 60), and a sandboxed `context_refresh` against the real clone seeds all seven cleanly, frontmatter stripped, no-op on unchanged HEAD, whole-swap refusal + cache-hold verified against a deliberately broken master. - **Deployment risk 2 (all deployments unseeded at merge)**: covered by the same evidence -- the production first-tick seed will succeed against current master. I will verify the LIVE seed on the host right after merge and confirm surfaces source before I stand down. - **Clone-path divergence**: moot -- `AGENT_REPO_ROOT="$AGENT_STATE_DIR/repos"` is hardcoded (tick.sh:26), so both paths derive from the same root. - **check-sync now guards a document the model no longer reads**: CONFIRMED and the sharpest catch of the round. Bounded tonight (igor#483 made AGENTS.md and the sourced worker-contract byte-identical), but drift begins with the next brain-side edit. Follow-up ticket filed covering the retarget plus this round's nits (atomicity comment overclaim, install.sh env-source placement, test-feedback isolation comment, negative wiring assertion). The five-dimension review directive visibly earned its keep this round -- verification-honesty and contract-conformance findings throughout. Merging on green.

igor added 1 commit 2026-08-09 06:27:07 +00:00

Merge branch 'master' into agent/485-feat-source-prompt-surfaces-from-the-distillery c6988c8aa8

igor merged commit 8b715ca7b1 into master 2026-08-09 06:27:29 +00:00

igor deleted branch agent/485-feat-source-prompt-surfaces-from-the-distillery 2026-08-09 06:27:29 +00:00

igor referenced this pull request from a commit 2026-08-09 06:27:30 +00:00

Merge pull request 'feat: source igor's prompt surfaces from the Distillery at master, with a last-good cache (no in-repo fallback)' (#486) from agent/485-feat-source-prompt-surfaces-from-the-distillery into master

igor referenced this pull request 2026-08-09 06:27:56 +00:00

chore: retarget check-sync to the sourced worker contract; delete unconsumed in-repo prompt copies #487

igor referenced this pull request 2026-08-09 07:02:09 +00:00

chore: delete unconsumed in-repo prompt copies (operator decision: guard interaction) #489

igor referenced this pull request from a commit 2026-08-09 07:14:39 +00:00

WIP: issue #487 checkpoint -- revert out-of-scope deletions per rescope

igor referenced this pull request 2026-08-09 07:17:14 +00:00

chore: retarget check-sync's sentinel gate to the worker's real system prompt #488

igor referenced this pull request 2026-08-09 07:29:37 +00:00

chore: retarget check-sync's sentinel gate to the worker's real system prompt #488

Sign in to join this conversation.

Reviewers

No reviewers

joshtronic

Labels

Clear labels   

Agent

  

Compat/Breaking

Breaking change that won't be backward compatible

  

Kind/Bug

Something is not working

  

Kind/Documentation

Documentation changes

  

Kind/Enhancement

Improve existing functionality

  

Kind/Feature

New functionality

  

Kind/Security

This is security issue

  

Kind/Testing

Issue or pull request related to testing

  

Priority

Critical

The priority is critical

  

Priority

High

The priority is high

  

Priority

Low

The priority is low

  

Priority

Medium

The priority is medium

  

Reviewed

Confirmed

Issue has been confirmed

  

Reviewed

Duplicate

This issue or pull request already exists

  

Reviewed

Invalid

Invalid issue

  

Reviewed

Won't Fix

This issue won't be fixed

  

Status

Abandoned

Somebody has started to work on this but abandoned work

  

Status

Blocked

Something is blocking this issue or pull request

  

Status

Need More Info

Feedback is required to reproduce issue or to continue work

No labels

Agent

Compat/Breaking

Kind/Bug

Kind/Documentation

Kind/Enhancement

Kind/Feature

Kind/Security

Kind/Testing

Priority

Critical

Priority

High

Priority

Low

Priority

Medium

Reviewed

Confirmed

Reviewed

Duplicate

Reviewed

Invalid

Reviewed

Won't Fix

Status

Abandoned

Status

Blocked

Status

Need More Info

Milestone

Clear milestone

No items

No milestone

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Total time spent: 2 minutes 44 seconds

igor

2 minutes 44 seconds

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

joshtronic/igor!486

No description provided.