agentsclimarketplace

Ops

Skill HDHNezherParking-cum-Y638-Intl-Ltd/titan/skills/ops

A disciplined 10-stage autonomous agent pipeline (Reflexion, Tree Search, ReAct) for Claude Code and AI coding agents. Lifts local LLMs to senior engineering quality.

Install
npx -y skills add HDHNezherParking-cum-Y638-Intl-Ltd/titan --skill ops

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 27 days oldThe repository was created 27 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Session operations spine — one continuous workflow: PREFLIGHT (verify repo state before trusting anything) → work → LAND (gate, commit, ledger, lessons, staleness sweep), plus DECIDE (owner-decision interview → dated record → propagation sweep) and COMMISSION (external template/prompt intake → conflict adjudication by the record). Use at session start ('preflight', 'resume', 'where were we'), when finishing a chunk ('land this', 'ship and log', 'wrap up'), when an owner-only decision is needed ('decide X', 'founder decision', 'interview me'), or when an outside prompt/template/brief arrives ('audit this commission', 'a template says...').

SKILL.md

5.0 KB, as published. Nobody here has run it

ops — session operations spine

Four modes, one continuous loop: PREFLIGHT → work → LAND, with DECIDE and COMMISSION as event modes inside any session. Auto-detect mode from context; explicit arg wins. Generic in any repo. If project files exist — CLAUDE.md, root MEMORY.md, docs-requiring-human-input.md, vault/lessons_learned.md, decision-record templates (docs/ops/*template*) — use them. Never invent structure a repo doesn't have.

PREFLIGHT (session start / resume — every session, including your own resumed ones)

  1. git log --oneline -10 + git status.
  2. Read the newest entry of the root ledger (MEMORY.md or equivalent) + project CLAUDE.md authority/landmine sections.
  3. Carried-over claims = hypotheses. Re-verify each load-bearing one against code/commits before acting. Docs drift; audits produce false alarms; your own prior session counts as carried-over.
  4. If the project defines a gate (npm run check or similar): run it BARE, record counts with today's date as the baseline.
  5. Output a 5-line state brief: HEAD · ledger-newest · drift found (fixed or flagged) · gate baseline · next queued item.

LAND (finish a work chunk — never skip, docs-only included)

  1. Gate BARE, exit 0. Piping allowed only with real exit capture (; echo EXIT:${PIPESTATUS[0]}). Red gate → fix first; never land red, never claim a skipped check ran.
  2. Staleness sweep on every register the chunk touched: re-derive counts fresh, stamp "as of YYYY-MM-DD", sweep owner-action queues for rows this work resolved (queue rows rot when work ships).
  3. Ledger: append/update the root ledger, newest-first — what shipped, commit hashes, gate counts, owner-queue deltas.
  4. Lesson: if something transferable was learned, ONE line appended to the project lessons file (single file, never forked). Nothing learned → skip; don't invent a lesson.
  5. Commit per task: message = what + why; include the host's co-author trailer if it defines one. NEVER push unless asked.
  6. Output: landed summary — commits, gate proof, ledger/lesson lines written, queue deltas.

DECIDE (owner-only decision needed)

  1. Confirm it IS owner-only (project CLAUDE.md lists them). Mechanical question with one right answer → decide yourself, note reasoning, no interview.
  2. Assemble what the record already constrains: prior commitments, hard guardrails, risk register. If a prior adjudication covers it → re-apply by citation; a repeat assertion does not re-open a decided question.
  3. ONE batched AskUserQuestion: options with cost / risk / reversibility, recommendation first and marked; previews showing the resulting doc lines when that helps comparison.
  4. Record it: use the project's decision-record template if one exists (e.g. docs/ops/*template*human-review/decision-<slug>-<YYYYMMDD>.md); otherwise minimal record: question · constraints · options · DECISION (owner's words) · propagation list · supersession note.
  5. Propagate: grep the topic across the repo; amend EVERY doc that assumed otherwise; list each in the record. A decision that isn't propagated is a future false claim.
  6. Finish with LAND.

COMMISSION (external template / prompt / brief arrives)

  1. Census its factual claims first (counts, file lists, corpus size) against reality — commission prompts misstate corpora; report real numbers.
  2. Diff its directives against the ratified record (project CLAUDE.md → amendments → canonical plan → risk register, in that precedence). Conflicts → resolve BY the record, assign conflict IDs in the project's ledger pattern, and re-refuse repeats by citation.
  3. Execute only the aligned remainder. Refused/deferred items → owner-review row with citations, not silent drops.
  4. Finish with LAND.

Integration (the continuous workflow)

  • titan skill (this plugin): PREFLIGHT runs before pipeline stage 1 and feeds the Recall stage; LAND is the post-Gate completion record; DECIDE is the "user challenge" escalation path; COMMISSION wraps intake whenever the request is an outside brief.
  • Spec-pipeline plugins (multi-day features): the spec pipeline owns task state; PREFLIGHT/LAND still bookend each execution session.
  • Always-binding rules: zero fabrication (counts re-derived + dated) · underclaim > overclaim · ask before deleting · no unilateral owner decisions · inline judgment passes over cold spawns (one context-inheriting fork max) · native tools before shell.

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.