Shiplog

Shiplog

The captain's log for your codebase. Every decision, discovery, and change logged as you ship code.

Use GitHub as a complete knowledge graph where every brainstorm, commit, review, and decision is traceable. This skill orchestrates existing skills and references; it defines when and how to invoke them and what documentation protocol to follow.

Core Principle

Nothing gets lost. Every brainstorm becomes an issue. Every issue drives a branch. Every branch produces a PR. Every PR is a timeline of the entire journey. Git becomes the uber-memory.

Golden-Path Walkthrough

A complete issue-to-merge example using a concrete fake issue #999. Follow this sequence and every artifact produced will pass the acceptance checklists in commands/shiplog/*.md.

Step 1 — Plan Capture (/shiplog plan)

A brainstorm produces: "Add rate-limit headers to the API so clients can back off intelligently." Load commands/shiplog/plan.md and create the issue:

gh issue create \
  --title "[shiplog/plan] Add rate-limit headers to API" \
  --label "shiplog/plan" \
  --body-file .tmp-plan.md

Issue body (excerpt):

<!-- shiplog:
kind: state
status: open
phase: 1
readiness: ready
task_count: 2
tasks_complete: 0
max_tier: tier-2
updated_at: 2026-04-18T10:00:00Z
-->
## Tasks
- [ ] **T1: Add X-RateLimit-* headers to response middleware** `[tier-2]`
  - **What:** Inject X-RateLimit-Limit, X-RateLimit-Remaining, X-RateLimit-Reset on every API response.
  - **Verification:** curl the endpoint; confirm all three headers are present.
- [ ] **T2: Document header semantics in API reference** `[tier-2]`
  - **What:** Add a "Rate Limiting" section to docs/api-reference.md.
  - **Verification:** Section exists and matches header values in integration test.

Authored-by: claude/sonnet-4.6 (claude-code)

Apply lifecycle label: gh issue edit 999 --add-label "shiplog/ready"

Step 2 — Branch Setup (/shiplog start)

Load commands/shiplog/start.md. Create the branch from the default branch tip (substitute your repo's default branch, e.g. main or master):

git fetch origin <default-branch>
git checkout -b issue/999-rate-limit-headers origin/<default-branch>

Swap label and update envelope:

gh issue edit 999 --remove-label "shiplog/ready" --add-label "shiplog/in-progress"
# Edit issue body: readiness: ready → readiness: in-progress

Post session-start comment:

[shiplog/session-start] #999: Starting work

**Branch:** `issue/999-rate-limit-headers`
**Starting tasks:** T1, T2
**Plan:** Add headers in middleware (T1) then document (T2).

Authored-by: claude/sonnet-4.6 (claude-code)

Step 3 — Commit Context (/shiplog commit)

After implementing T1, load commands/shiplog/commit.md and commit:

git add src/middleware/rate_limit.ts
git commit -F .tmp-commit-msg.md
# Commit message: feat(#999/T1): add X-RateLimit-* headers to response middleware
#
# Injects the three standard headers into every API response via middleware.
#
# Authored-by: claude/sonnet-4.6 (claude-code)

Post commit-note on the issue:

[shiplog/commit-note] #999: a1b2c3d feat(#999/T1): add X-RateLimit-* headers

**What:** Middleware now injects X-RateLimit-Limit, -Remaining, and -Reset.
**Why:** Clients need these to implement back-off without polling.
**Verification:** curl localhost:3000/api/health — all three headers present.

Authored-by: claude/sonnet-4.6 (claude-code)

Repeat for T2 commit (docs(#999/T2): document rate-limit headers in API reference).

Step 4 — PR Timeline (/shiplog pr)

Load commands/shiplog/pr.md. Push and create the PR:

git push -u origin issue/999-rate-limit-headers
gh pr create \
  --title "feat(#999): add rate-limit headers to API" \
  --label "shiplog/history" \
  --label "shiplog/issue-driven" \
  --body-file .tmp-pr-body.md

PR body includes: envelope (kind: history), summary, Closes #999, Journey Timeline with both commits, Verification checklist, Reviews placeholder, and sig block:

Authored-by: claude/sonnet-4.6 (claude-code)
Last-code-by: claude/sonnet-4.6 (claude-code)
*Captain's log — PR timeline by **shiplog***

Transition issue label:

gh issue edit 999 --remove-label "shiplog/in-progress" --add-label "shiplog/needs-review"

Step 5 — Cross-Model Review (/shiplog review)

A different model (e.g., openai/gpt-5.4 (codex)) loads commands/shiplog/review.md, reads the diff, and posts the signed review comment on the PR:

[shiplog/review-handoff] #999: Review of PR #<PR>

Middleware approach is correct. Headers match RFC 6585 semantics.

Reviewed-by: openai/gpt-5.4 (codex, effort: high)
Disposition: approve
Scope: full diff
Follow-ups: none

The reviewer then updates the PR body review snapshot in place:

Current state: approved
Last reviewed by: openai/gpt-5.4 (codex, effort: high)
Last reviewed at: 2026-04-18T15:00:00Z
Reviewed commit: a1b2c3d
Source artifact: <URL to review comment>
Needs re-review since: —

Step 6 — Merge

Cross-model gate is satisfied (Last-code-by: claude/sonnet-4.6 ≠ Reviewed-by: openai/gpt-5.4). Merge the PR. Remove all lifecycle labels from issue #999. GitHub closes #999 automatically via Closes #999.

When This Skill Activates

User-invocable: /shiplog, /shiplog models, /shiplog <phase>

/shiplog models: Re-runs the routing setup prompt. See references/model-routing.md.

Auto-activate when ANY of these occur:

• User says "let's plan", "let's brainstorm", or "let's design"
• User explicitly requests traceability or knowledge-graph tracking
• Creating a new issue or PR with intent to document decisions
• Mid-work discovery requiring a new issue or stacked PR
• User asks "where did we decide X?" or "what's the status of Y?"
• Resuming work on an existing issue or PR
• Applying review feedback, fixing review findings, or addressing request-changes dispositions
• User references an issue or PR by number
• Issue or PR body footer contains Managed by **shiplog** (recognition trigger — do not add this footer proactively)

Do NOT auto-activate for:

• Generic coding requests that do not need traceability
• Simple bug fixes or refactors that do not need durable workflow history
• Work where a more specific skill is the better fit

Verb Grid

Each /shiplog <phase> slash command maps to one sub-skill file. Load the file for that phase — it owns policy, templates, and acceptance checklist.

/shiplog models: Re-runs the model-routing setup prompt. See references/model-routing.md.

Canonical Kind → Tag → Label Map

kind: in envelope YAML is the source of truth. Title tags ([shiplog/<tag>]) and GitHub labels (shiplog/<label>) are derived views on kind:. When a new artifact kind is defined, it is added here once and all three surfaces are updated together.

Lifecycle labels (shiplog/ready, shiplog/in-progress, shiplog/needs-review) are not tied to an artifact kind:. They track issue/PR workflow state and are mutually exclusive. Apply them per the Triage Field Maintenance table below.

Aspect labels (shiplog/discovery, shiplog/stacked, shiplog/issue-driven) classify the relationship between artifacts, not their content kind. They may coexist with lifecycle labels.

For the full label set, color codes, and bootstrap CLI snippets, see references/labels.md (which cross-references this table as its canonical kind source).

ID-First Naming Convention

All artifacts use #ID as the primary key for fast, token-efficient retrieval.

Semantic tag vocabulary for user-facing headings: plan, session-start, session-resume, commit-note, discovery, blocker, implementation-issue, handoff, review-handoff, history, amendment, milestone, and verification. Format: [shiplog/<tag>] <human title>. See the Canonical Kind → Tag → Label Map section for the authoritative mapping.

Task IDs: Tasks carry local IDs (T1, T2, ...) scoped to the issue. Commits use #<id>/<Tn>.

Retrieval: gh issue list --search "#42" | git log --grep="#42" | git log --grep="#42/T1" | gh pr list --search "#42"

Envelope Schema and Triage Fields

Issue and PR envelopes use <!-- shiplog: ... --> HTML comment blocks. The full field schema and triage field derivation rule live in references/artifact-envelopes.md §1. Summary of triage fields:

Triage field maintenance: keep envelope triage fields current on each lifecycle event. See the Triage Field Maintenance table below. Derivation rule is in references/artifact-envelopes.md §1.

Triage Field Maintenance

Issue envelope triage fields (readiness, task_count, tasks_complete, max_tier) and lifecycle labels must be kept current so triage scans produce accurate results.

Derived fields: task_count, tasks_complete, and max_tier are computed from the issue body task list (counted - [ ] / - [x] lines and highest [tier-N] among unchecked tasks). See the derivation rule in references/artifact-envelopes.md §1 "Triage field derivation". Only readiness is hand-written.

Edit the issue body in place when these fields change. Triage metadata is derived state, so refreshing it does not require Updated-by: provenance.

Agent Identity Signing

Every shiplog artifact (comments, PR bodies, review sign-offs) must carry a provenance signature in the canonical format:

<role>: <family>/<version> (<tool>[, <qualifier>])

Signature field reference

Searching: Authored-by: → original authorship. Updated-by: → later material editors. Reviewed-by: → review artifacts. Last-code-by: → most recent code author on a PR branch. claude/ → all Claude artifacts. (codex → all Codex artifacts.

Model detection per tool

Orchestration role qualifiers

When an artifact is emitted as part of a multi-lane flow, qualifiers carry orchestration role information:

• orchestrator — the current actor dispatched or collected delegated lanes
• sub-agent: reviewer — delegated reviewer lane
• sub-agent: verifier — delegated closure verifier lane
• sub-agent: implementation — delegated implementation lane

Examples:

• Authored-by: openai/gpt-5.4 (codex, effort: high; orchestrator)
• Reviewed-by: claude/opus-4.6 (claude-code, sub-agent: reviewer)
• Authored-by: openai/gpt-5.4 (codex, sub-agent: verifier)

Correction rule

If a shiplog artifact carries an incorrect or incomplete signature, correct it in place when the platform allows editing. Otherwise post an immediate follow-up correction.

Edit provenance rules

• Authored-by: records the original author of an artifact body.
• Updated-by: records a later model or human who materially edits that same artifact body. Preserve the original Authored-by: line and append a new Updated-by: line for each material edit, newest last.
• Reviewed-by: is review-only. Do not use it for authorship or edit attribution.
• Updating a PR body's review snapshot after publishing a signed review comment or after pushing code that makes a prior review stale counts as a material edit.
• A material edit changes meaning, facts, scope, requirements, acceptance criteria, verification results, review disposition, or a handoff contract. Typos, formatting cleanups, and link-only fixes are cosmetic and do not need Updated-by:.

Code provenance: Last-code-by:

Last-code-by: tracks which model most recently pushed code to a PR branch. It is distinct from artifact provenance fields.

When to set Last-code-by::

• On PR creation: set in the PR body sign-off block (the creating model is the initial code author).
• After pushing code to an existing PR branch: update via gh pr edit.
• After review-driven code changes: the reviewer who pushes fixes becomes Last-code-by:.

When NOT to update Last-code-by:: reviewing without pushing code; editing the PR body text without commits; rebasing or force-pushing without new code changes.

Why this field exists: The multi-model review gate must know who last changed the code to determine whether a review is cross-model (gate-satisfying) or same-model (non-gate-satisfying). Without Last-code-by:, consumers fall back to git commit forensics.

Fallback chain for review gating:

1. Last-code-by: in the PR body (authoritative)
2. Updated-by: in the PR body (approximate)
3. Authored-by: in the PR body (original author — may be stale)
4. Git commit author on the PR branch (last resort)

Edit-in-place vs amendment

• Edit in place when the artifact is meant to stay the single canonical current body: issue bodies, PR bodies, and latest-wins status/history artifacts. Refresh envelope updated_at and add updated_by plus edit_kind fields when an envelope exists.
• Post an amendment artifact when the original text matters for auditability: handoffs, verification comments, commit-note comments, review sign-offs, and other major signed timeline entries.

Use supersedes when the new artifact replaces the old one as canonical. Use amends when the new artifact corrects or clarifies but both should remain visible.

In-place edit footer — append after original Authored-by::

Updated-by: <family>/<version> (<tool>)
Edit-kind: correction | amendment | rewrite
Edit-note: [1 sentence describing what changed and why]

Amendment artifact template:

<!-- shiplog:
kind: amendment
issue: <ISSUE_NUMBER>
pr: <PR_NUMBER>
updated_at: <ISO_TIMESTAMP>
amends: <artifact-reference>
-->

## [shiplog/amendment] #<ISSUE_NUMBER>: <brief description>

**Target:** [URL to the artifact being corrected or clarified]
**Edit kind:** correction | amendment | rewrite
**Why new artifact:** [why this should not be a silent in-place edit]
**What changed:**
- [change 1]
- [change 2]

**Current canonical artifact:** [URL to the current body, or `this comment`]

Authored-by: <family>/<version> (<tool>)

If the amendment fully replaces the old artifact, swap amends: for supersedes: and update the old artifact with superseded_by: when practical.

PR body review snapshot maintenance

When a PR body carries the current review snapshot:

• Post the signed Reviewed-by: comment first. That comment is the review evidence.
• Then refresh the PR body snapshot in place so retrieval flows can read current review state without replaying the comment thread.
• If new code lands after the latest signed review, update the snapshot to needs-rereview and record the commit that made the prior review stale.
• Use the standard Updated-by: footer and envelope updated_by / edit_kind fields for these edits.

Model identity detection is also used by model-tier routing to verify the current model matches the recommended tier. See references/model-routing.md.

GitHub Labels

shiplog manages a compact repo-level label vocabulary so issues and PRs stay filterable even when a reader never opens the body. See references/labels.md for the canonical label set, descriptions, and CLI snippets.

Label rules:

• On the first write operation in a repo, bootstrap or refresh labels with gh label create --force ....
• Apply labels at creation time with gh issue create --label or gh pr create --label.
• shiplog/blocker is stateful. Add it when work becomes blocked and remove it when the blocker is cleared.
• shiplog/ready, shiplog/in-progress, and shiplog/needs-review are mutually exclusive lifecycle labels.

Mandatory Issue Capture

Implementation trouble that materially affects the work must be durably recorded before the agent proceeds to the next material step or ends the turn.

What counts as a relevant implementation issue

• Failed attempts
• Hidden dependencies
• Risky workarounds
• Scope surprises
• Verification gaps
• Environment or tooling friction

What does not require capture

• Normal iteration where the final approach is obvious from the diff
• Minor typos or lint fixes resolved in the same commit
• Expected complexity that matches the task description

Capture rule

The timeline comment is the minimum: one paragraph explaining what happened, why it matters, and how it was resolved or deferred.

Edge Cases

• No issue exists: Let the user work. At first commit or PR, offer to create a tracking issue.
• Mid-work activation: Check branch name for issue/N-*. If found, add catch-up timeline comment via shiplog:timeline. If not, offer retroactive issue creation.
• Small tasks (< 30 min): Lightweight protocol - issue optional, branch still created, PR sections can be brief.
• Hotfix / emergency: Fix first. Create issue and PR after, backfilling the timeline.
• Post-merge cleanup: Remove a worktree only when its branch is merged, no open PR still depends on it, and it is not the active workspace. See references/orchestrator-protocol.md.

Requirements

All recommended skills are optional. The current optional integrations are listed below. Without them, shiplog falls back to direct gh/git commands.

User-Facing Language

The phase numbers are internal workflow labels. Do not surface them to the user.

Preferred labels: Plan Capture, Branch Setup, Discovery Handling, Commit Context, PR Timeline, History Lookup, Timeline Updates.

Brand formatting: Always bold the word shiplog in user-facing text (messages, comments, PR bodies, issue bodies). Write it lowercase and bold: shiplog. This does not apply to code identifiers, branch names, CLI output, or other machine-readable contexts where markdown is not rendered.

Shell Portability

Keep the workflow cross-platform. See references/shell-portability.md for full guidance and Bash/PowerShell patterns.

Key rules:

• Prefer gh ... --body-file <temp-file> for multiline content.
• Break chained shell commands into separate steps when the shell operator differs.
• Keep Bash examples as the primary path; add PowerShell notes where syntax diverges.

Integration Map

This skill ORCHESTRATES. Sub-skills under commands/shiplog/ each own their phase's policy, runnable queries, and acceptance checklist in a single file — no cross-file navigation required during execution. References in references/ are deep-dive anchors for cross-cutting policy only.

Sub-skill map

Cross-cutting reference anchors

External skill delegation

Graceful degradation: Co-located sub-skill → references deep-dive → external skill → direct gh/git commands. Minimum viable installation: gh CLI + git + this skill.

Conflict avoidance: This skill sets the WORKFLOW context. External skills provide IMPLEMENTATION helpers. Shiplog's internalized conventions always take precedence for artifact format, signing, labels, and review gates.

Runtime-Aware Orchestration

Shiplog records orchestration honestly instead of assuming one agent backend fits every runtime.

• Local parallel tool fan-out: one orchestrator runs multiple independent helper calls in parallel. Good for sidecar reads; not a separate reviewer identity.
• Bounded sub-agent: the orchestrator spawns a child lane with a scoped contract and collects a return artifact.
• External session delegation: a separate tmux session, terminal agent, or other durable worker runs the lane outside the current orchestrator.
• Contract-only fallback: shiplog emits the handoff or review contract when the current runtime cannot execute the lane itself.

Isolation backend is tracked separately from the orchestration primitive. A git worktree, forked workspace, or tmux session may all isolate delegated work, but only the primary feature branch/worktree is shiplog's canonical branch record.

See references/orchestrator-protocol.md for the capability mapping, fan-out templates, and cleanup protocol.

Optional External Skills

These skills enhance shiplog but are not required. Shiplog's conventions take precedence when both are active.

References

One line per deep-dive file. Open only when the co-located sub-skill or inline section is insufficient.