upstream-refresh
Skill15 starsChanged 18 days agoApache-2.0View on GitHub
Use when grim's upstream knowledge may be stale — a harness shipped or changed a capability, watchlist rows past six months, vendor changelogs since the last check, a deep re-verify of one harness (claude, codex, copilot…), spec or forge claims, or expired research.
In this repository45 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .agents/research/hooks_vendor_reports/claude.mdCLAUDE.md
- CLAUDE.mdCLAUDE.md
- code-docs-cleanupSkill
- docs-instrumentSkill
- docs-planSkill
- docs-reviewSkill
- hex-architectSkill
- hex-coreSkill
- hex-discussSkill
- hex-executeSkill
- hex-finalizeSkill
- hex-initSkill
- hex-loopSkill
- hex-planSkill
- hex-retroSkill
- hex-reviewSkill
- nox-reviewSkill
- bugfixSkill
- builderSkill
- changelogSkill
- code-checkSkill
- code-docs-cleanupSkill
- commitSkill
- docs-instrumentSkill
- docs-planSkill
- docs-reviewSkill
- docsSkill
- finalizeSkill
- hex-architectSkill
- hex-coreSkill
- hex-discussSkill
- hex-executeSkill
- hex-finalizeSkill
- hex-initSkill
- hex-loopSkill
- hex-planSkill
- hex-retroSkill
- hex-reviewSkill
- meta-maintain-configSkill
- meta-validate-contextSkill
- nextSkill
- nox-reviewSkill
- qa-engineerSkill
- security-auditorSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 2 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 49
Never compute ages by hand:
- Very specific: line 68
Preflight: Not on main; honor hex-state.md.
- Very specific: line 70
Pick: every non-noop target from the ladder.
- Distinctive: line 90
No per-target opus gate, no per-target verify, no ordering of research — only the ledger writes and code changes after collection serialize.
- Very specific: line 139
Every changed or newly dated ledger row carries verified YYYY-MM-DD, a primary-source URL, and the vendor version when known.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/upstream-refresh — keep grim's upstream facts current
Re-checks the dated upstream claims grim depends on — harness capabilities,
metadata keys, config paths and env vars, specs, forge APIs, the product
landscape, research expiry. Check ledger:
.agents/upstream-checks.md; per-domain
registry: references/domains.md. Design record:
harness-capability-freshness.md.
Modes
/upstream-refresh [<harness>] [--domain <domain>] [--force] [--focus "<text>"]| Invocation | Targets | Depth |
/upstream-refresh | every non-noop ledger row | per ladder, never deep |
/upstream-refresh <harness> | that harness row | deep (the only way to get one) |
--domain vendors | the 18 harness rows | per ladder, never deep |
--domain <catalog|specs|forge|landscape|research> | that one domain row | per ladder |
--force | as selected | a noop becomes sweep |
--focus "<surface>" | every harness row, or only the named harness | that surface only, ladder ignored |
Validate first. The positional token must be a harness name from the
ledger (claude, opencode, copilot, codex, cursor, kiro, junie,
gemini, zed, amp, antigravity, cline, droid, goose, warp,
openclaw, kilo, qoder); --domain must be vendors or a domain row
name. Anything else → stop before any edit, print the valid names. An
unknown token is never read as a focus.
Depth ladder
Never compute ages by hand:
python3 .claude/skills/upstream-refresh/scripts/upstream_stale.py ladder \
[--name <harness>] [--domain <domain>] [--force]It prints <target> <noop|feed|sweep|deep> <reason> per target. Exit 2 →
relay its stderr and stop. 0 days → noop; 1–2 days → feed (read the
change feed from the cursor); older or never checked → sweep; a named
harness → deep. A reason ending deep due → name that harness in the
report as due for /upstream-refresh <harness>. Every target
noop → print every target checked today — no-op (use --force), edit
nothing. task upstream:stale (the script's report) is the overview.
Default run
Dead simple, one sitting (1–2 hours):
- Preflight. Not on
main; honorhex-state.md. Read the ledger and domains.md. - Pick every non-
nooptarget from the ladder. - Fan out — everything at once. One sonnet update-check subagent per
target: every Tier 1 and Tier 2 harness and every domain in the pick,
spawned together, no tier ordering, none waiting on another. Each is
briefed from
researcher-brief.md with its depth
(
feed→ Feed procedure;sweep→ every ledger row and code stamp the target owns).catalogis confirm-only andresearchfollows its dating rule (domains.md). A silent subagent is retried once, then its row stays unchanged with a note. - Collect. Findings into
## Claims; drop any without a URL and a verbatim quote (list it under## Friction). - Write class (a) fixes (Routing): ledger dates with evidence, doc and
matrix date fixes. Update each completed target's ledger row —
Last check= today,Depth,Feed cursor(the feed's newest tag),Notes. A target that did not complete keeps its row and stays stale. - Report code-driving findings (renderer, metadata, validation) in the
run report and the artifact's
## Routing— they do not block the run. - Refine and commit (Refinement, Commits).
No per-target opus gate, no per-target verify, no ordering of research — only the ledger writes and code changes after collection serialize.
Landing a code change (only when this run implements one): one batched
opus verifier re-fetches the sources of every code-driving claim and signs
off or refutes in ## Verifier sign-off — refuted claims do not land. Then
implement per Routing (b) with tests and run one task verify before the
final commit.
Deep pass (named harness only): the default run plus deep-pass.md — full surface inventory and the sandboxed live CLI check.
Feed procedure (feed rung)
- Feed and cursor come from domains.md and the ledger row.
- GitHub feed — the paginated Releases API, never
releases.atom:gh api --paginate repos/<o>/<r>/releases --jq '.[] | [.tag_name, .prerelease] | @tsv', filtered by the tag stream domains.md names; entries above the cursor are new. Page feed (changelog) — headings above the cursor are new. - Cursor not in the fetched window → escalate to
sweep, record friction. Feed unreachable → keep the cursor, note it, go on. - Relevant entries touch a surface grim renders or documents (skills, rules, agents, MCP, hooks, frontmatter keys, config paths, env vars); others are skipped. Advance the cursor to the newest tag, heading or version seen.
Routing
| Class | What | How it lands |
| (a) | Ledger dates, docs text, matrix rows | Fixed directly, chore: or docs: commit |
| (b) | Additive renderer or metadata support | One commit: renderer + docs page + parity test + watchlist row, with tests |
| (c) | Compatibility; config-path or install-layout moves; new vendors; class-3 requests | Issue draft — never code |
| (c) security | Any security-relevant finding | (c) security — <file:line> · <class> in ## Routing only — no draft, no detail |
(b) rules (Principle 9, docs/src/content/docs/stability.md): additive
only; a self-heal test re-materializes an artifact installed before the
change and asserts grim status reports state == "installed". Both
src/install/client_target.rs parity tests must pass — red → revert and
draft an issue, never weaken the test. A docs-page change triggers the
catalog drift review (catalog/README.md). Hooks
(PR #98) are findings only.
Issues. gh issue create only under an issue-creation grant; otherwise
an issue draft (title, body, labels) under ## Issue drafts, listed in the
final report. Security findings are never drafted as issues.
Evidence
Every changed or newly dated ledger row carries verified YYYY-MM-DD, a
primary-source URL, and the vendor version when known. Each target writes
.agents/research/research_upstream_<target>_<YYYYMMDD>.md,
**Expires:** = today + 6 months:
# Research: upstream refresh — <target> (<depth>)
## Metadata
**Date:** <YYYY-MM-DD> · **Domain:** <domain> · **Depth:** <ladder line>
**Triggered by:** /upstream-refresh <args>
**Expires:** <today + 6 months, YYYY-MM-DD>
## Claims <!-- researcher-brief.md output table -->
## Verifier sign-off <!-- only when code lands; else "none required" -->
## Live CLI check <!-- deep only: command · pass|fail|skip · reason -->
## Routing <!-- claim → a|b|c → commit or draft; security: (c) security — <file:line> · <class> -->
## Issue drafts <!-- or "none" -->
## Friction <!-- dropped claims, feed escalations, retries -->
## Skill changes <!-- summary + commit, or "none" -->Refinement
Friction with this skill is fixed in the same run: record it under
## Friction, fix .claude/skills/upstream-refresh/** in its own
chore(skills): commit, and append one line to the discussion's
## Hand-off record:
- <date> · <target> · <depth> · skill: <change summary | no change> · artifact: <path>.
Friction with any other skill or rule goes to the retro inbox only
(hex-state.md).
Commits
- Conventional Commits with
--signoffon the current branch — nevermain, never pushed.chore:for ledger, research and skill commits;feat:/fix:/docs:for renderer and docs changes. - A pure ledger/doc run gates on
task claude:verify(plustask docs:checkwhen a docs page changed); if the commit hook still demands a freshtask verify, run it once for the whole run. A run that lands code runscargo fmtand onetask verifybefore its final commit.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

