feat(work-items): add mini-SDLC pipeline-shape SSOT + wire work Step 5 to it

[kyle-sexton]

Summary

Stage 1 of the mini-SDLC umbrella (#513). The /work-items:work lane had no durable definition of the shape of the pipeline it runs per item — the lane catalog, the implementer ≠ reviewer ≠ verifier invariant, and the depth tiers lived only as evolving skill prose and per-issue plans, so the shape drifted and could not be scaled or reviewed in one place.

This PR adds the single source of truth for that stable policy and points the work skill at it:

• New plugins/work-items/reference/pipeline-shape.md owns the stable policy: the fixed lane set (explore → research → plan → devil's-advocate → implement → test → review → verify, with the re-anchor slot reserved), the "variation in depth, never in shape" principle, the role-separation invariant, and placeholder depth tiers carried as a plan field. It points at the return-payload contract (loops/skills: orchestrator context economy — subagent return-payload contracts + GitHub-externalized state #496) and the convention-gap protocol (repo/ops: .work gitignore hides the multi-machine control surface — track loop prompts + operating procedures, codify the state boundary #554) rather than restating them.
• work/SKILL.md Step 5 now dispatches against the SSOT for lane shape, additively — the existing instruction to follow the consuming project's own development workflow and domain rules is retained; the chain runs the shape within the consumer's workflow and rules, never in place of them. There was no inline lane catalog in the skill to remove (the "no duplicated lane prose" criterion is satisfied by absence), so this wire-in is purely additive.

Scope and honesty about current runtime

This stage lands the shape + wire-in only. The depth-scaling dispatcher and the separated-reviewer/verifier runtime are later #513 stages. The SSOT is written in normative (target-shape) voice and makes no claim that the work lane already depth-scales lanes or already fully separates the three roles today; rollout state is a single pointer to #513 rather than an inline stage-status section that would drift.

Nothing is locked: form, location, and name are a reversible reference-doc STOPGAP (OPERATOR DECISION 1 on #513, promotable to a skill later — naming-collision with session-flow:orchestrate / implementation:implement-dispatch flagged for that time), and the depth-tier names are placeholders (OPERATOR DECISION 2).

Fresh-docs verification

Per the repo fresh-docs mandate, confirmed this session from current official docs:

• A plugin's reference/*.md supporting files are auto-discovered and referenced by relative path from SKILL.md — no manifest registration needed. https://code.claude.com/docs/en/plugins-reference (Skills: "automatically discovered", "supporting files alongside SKILL.md"; "The manifest is optional... auto-discovers components"), https://code.claude.com/docs/en/skills ("detailed reference documentation... Reference these files from your SKILL.md").
• plugin.json version is an optional semver string; bump MINOR for new features and document in CHANGELOG.md. https://code.claude.com/docs/en/plugins-reference (version field; Version management). marketplace.json pins no version for work-items, so only plugin.json is bumped: 0.18.2 → 0.19.0 (rebased over a concurrent 0.18.2 patch that landed on the same plugin during review; reconciled so the 0.19.0 entry sits above 0.18.2 in the CHANGELOG and the minor bump stays the highest version).

Verification (local gates)

• check-changelog-parity.sh --check-bump origin/main — PASS (0.19.0 entry present)
• check-skill-portability.sh origin/main — PASS (no unexcused coupling tokens)
• check-changed-skills.sh origin/main — PASS (0 errors; 1 pre-existing no-Gotchas warning)
• validate-plugins.sh — PASS (all manifests + catalog)
• markdownlint-cli2 + typos on changed files — clean
• Independent fresh-context code review (producer ≠ reviewer) — changes-needed (minor) folded in: reworded four target-state-vs-present-tense spots so the SSOT defines the shape without claiming the runtime already depth-scales or fully separates roles, converted issue refs to the repo's #nnn code-span house style, fixed a re-anchor link text/target mismatch, and cited the settled sizing heuristic.

Closes #613

Related

• work-items(work): per-issue execution should compose installed skills as a role-separated mini-SDLC (implementer ≠ reviewer ≠ verifier, ephemeral fresh-context agents) #513 — mini-SDLC umbrella (parent); this is stage 1, and OPERATOR DECISIONs 1/2 govern the STOPGAP form and placeholder tier names
• loops/skills: orchestrator context economy — subagent return-payload contracts + GitHub-externalized state #496 — orchestrator context economy / subagent return-payload contract (pointed at, not restated)
• repo/ops: .work gitignore hides the multi-machine control surface — track loop prompts + operating procedures, codify the state boundary #554 — convention-gap protocol / interim-unblock template (pointed at, not restated)
• quality-net: skill changes ship with zero regression detection — 120 evals.json unexecuted, skill-quality:check uninvoked (staged eval/gate/pass^k umbrella) #530 — skill-regression eval umbrella (sibling prevention layer)

🤖 Generated with Claude Code

[claude]

Claude finished @kyle-sexton's task in 2m 35s —— View job

---

PR Review: feat(work-items): add mini-SDLC pipeline-shape SSOT + wire work Step 5 to it

• Gather context and classify request
• Read changed files
• Review pipeline-shape.md (new reference doc)
• Review work/SKILL.md (step 5 wire-in)
• Review plugin.json (version bump)
• Review CHANGELOG.md (new entry)
• Post findings

---

Summary

Clean, minimal Stage 1 of #513. No blockers. Two minor observations worth flagging for awareness, neither requiring changes before merge.

---

work/SKILL.md — Step 5 wire-in

The actual diff confirms the change is purely additive: one sentence inserted between dispatch is only *how* that workflow is carried out. and the #572 deferral note. No prior constraint language was removed. The ${CLAUDE_PLUGIN_ROOT} path reference is correct for plugin-form safety. ✅

The consumer-workflow primacy ("the dispatched chain runs that shape within the consumer's workflow and rules, never in place of them") lands cleanly alongside the existing MUST instruction. ✅

---

reference/pipeline-shape.md — new SSOT

What works well:

• Target-state voice is consistently maintained throughout ("is being built to run", "the runtime already…makes no claim"). The doc defines policy without overstating runtime behavior. ✅
• STOPGAP status, operator decision markers (1, 2), and the reversibility note are prominent and correctly placed. ✅
• Issue refs all use the `#NNN` code-span house style. ✅
• "Variation in depth, never in shape" principle + separation invariant are clearly owned here and nowhere else. ✅
• Placeholder depth tier names in angle brackets signal their provisional status unambiguously. ✅
• The "Contracts this shape composes" section correctly points at #496 and #554 rather than restating them. ✅

Observation 1 — Lane catalog references uninstalled skill names (expected, but worth recording)

The table at line 34–43 cites skill compositions including review:code-reviewer, review:quality-gate, verification:*, re-anchor:*, and tdd:principles that may not exist in this marketplace yet. The doc calls this out correctly ("the skills named are the intended realization of each lane, not a re-specification of their contracts"), so this is intentional — but it means a reader following these names today may find dead ends. Suggest adding a sentence to the lane catalog intro along the lines of "Not all named skills are installed in this marketplace today; the catalog defines intended realization, and #513 stages track their rollout." Or leave it to the STOPGAP/#513 pointer already present — either is defensible.

Observation 2 — Sizing heuristic cross-reference is opaque to future readers

Line 69: "matching the sizing heuristic settled in the #513 ACCEPTED decision brief" gives readers no inline anchor for what the heuristic actually is. This is a deliberate "point, don't restate" tradeoff (noted in the PR description), but if the heuristic can be named in a few words (e.g., "complexity × blast radius tiers") it would make this sentence navigable without opening #513. Not required — the pointer is sufficient given the explicit STOPGAP form.

---

plugin.json — version bump

0.18.2 → 0.19.0: correct MINOR bump for new feature (new reference file + SKILL.md pointer). ✅

Side note: the commit message body reads "Per-plugin bump 0.18.1 -> 0.19.0" — the from-version is stale (actual prior version was 0.18.2, per the CHANGELOG and the diff). The files are correct; the commit message line is cosmetically off. The PR description acknowledged this reconciliation explicitly, so it's recorded. Not a merge blocker.

---

CHANGELOG.md

The 0.19.0 entry is well-structured: Added + Changed sections, scope note at the end of the Added bullet, consistent format with prior entries. The scope note ("this stage lands the shape and the wire-in only…") correctly sets expectations without burying the lede. ✅

---

CLAUDE.md / design-rule alignment

Rule	Status
Repo-agnostic (no hardcoded paths)	✅
${CLAUDE_PLUGIN_ROOT} for within-plugin refs	✅
Configurable without editing the plugin	✅ (reference doc, no new userConfig needed)
Plugin-form-safe (no ../ reach-outs)	✅
No PII / secrets	✅
Versioned with explicit semver	✅
Security review (no code execution, no egress, no remote MCP)	✅

---

No blockers. The two observations are informational; Observation 1 could optionally be addressed before merge but is not required.

[claude]

Claude finished @kyle-sexton's task in 52s —— View job

---

Security Review

• Gather PR diff and changed files
• Review plugin.json changes
• Review pipeline-shape.md (new reference doc)
• Review SKILL.md change
• Review CHANGELOG.md
• Assess for injection surfaces, secrets exposure, and dangerous patterns
• Post security findings

---

No security issues found.

All four changed files are documentation and policy content. Here is what was checked:

plugin.json (+1/-1) — a semver version bump from 0.18.2 → 0.19.0. No new permissions, no new userConfig keys, no new MCP server references, no egress changes. Existing userConfig entries (work_dispatch_concurrency_cap, work_cycle_batch_cap) are unchanged.

reference/pipeline-shape.md (+80/-0) — a new normative reference document describing the desired pipeline lane catalog and role-separation invariant. It contains no shell expansions, no templated variable substitutions, no command execution, and no external URLs. Issue references (#496, #513, #513, #554) are inert GitHub fragment links. No injection surface and no credential or secret exposure.

skills/work/SKILL.md (+1/-1) — one sentence added to Step 5 pointing ${CLAUDE_PLUGIN_ROOT}/reference/pipeline-shape.md as the lane-shape authority. The ${CLAUDE_PLUGIN_ROOT} interpolation is the plugin's standard read-only path variable already used throughout this file; no new expansion or execution path is introduced. The change is purely additive prose.

CHANGELOG.md (+26/-0) — changelog entries only. No security relevance.

No code execution paths, secrets handling, permission changes, dangerous GitHub Actions patterns, injection surfaces, or supply-chain risk were introduced by this PR.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6886035f8a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[claude]

Claude finished @kyle-sexton's task in 1m 45s —— View job

---

PR Review: feat(work-items): add mini-SDLC pipeline-shape SSOT + wire work Step 5 to it

• Gather context and classify request
• Read git diff and changed files
• Review latest commit (standalone-install lane fallback)
• Review pipeline-shape.md
• Review work/SKILL.md
• Review plugin.json and CHANGELOG.md
• Post findings

---

Summary

Two commits. The original feature (6886035) was reviewed comprehensively in the prior pass — no new concerns there. This review focuses on the synchronize trigger commit (91c0f89) and confirms the full PR is clean.

---

Latest commit — 91c0f89: standalone-install lane fallback

This commit addresses the Codex P1 finding ("Add fallbacks for independently installed plugins"). Two targeted edits to pipeline-shape.md:

Principle section (lines 24–27)
The realization collapse spectrum extends from two levels to three, adding "the consumer's own workflow step for that stage when no lane skill is installed" as the bottom rung. This is well-placed — it sits inside the existing parenthetical that already described collapse levels, so the structure is preserved. ✅

Lane catalog intro (lines 31–35)
Explicitly states work-items installs standalone, hard-depends on none of the named skills, and specifies the fallback: "degrade to the consumer's own workflow for that stage (the Step 5 anchor) rather than dead-end." Directly answers the dead-end concern. ✅

Scope check:

• Target-state voice maintained throughout: "is to degrade" (not "degrades"). ✅
• No hard dependency declared — correctly deferred to #513/dispatcher stage. ✅
• The "Step 5 anchor" cross-reference is effective: SKILL.md Step 5 already mandates following the consumer's development workflow, so this ties the policy doc back to the existing runtime anchor without restating it. ✅
• The Codex concern asked for per-lane conditionality or dependency declarations. The fallback is stated uniformly (applies to all lanes) rather than per-row in the table — that's correct, since the fallback mechanism is identical for every lane. A per-row repetition would add noise without clarity. ✅

One minor observation: The second commit touches only pipeline-shape.md; the 0.19.0 CHANGELOG entry is not updated to mention the fallback clarification. This is defensible — the fallback is a within-release refinement of the existing 0.19.0 feature, not a new addition. The entry's summary still accurately describes what shipped. Not a blocker.

---

plugin.json — version

0.19.0: correct MINOR bump for a new reference file + SKILL.md pointer. No new permissions, MCP servers, or userConfig keys. ✅

---

CLAUDE.md design-rule alignment

Rule	Status
Repo-agnostic (no hardcoded paths)	✅
${CLAUDE_PLUGIN_ROOT} for within-plugin refs	✅
Configurable without editing the plugin	✅
Plugin-form-safe (no ../ reach-outs)	✅
No PII / secrets	✅
Versioned with explicit semver	✅
Security (no code execution, no egress, no remote MCP)	✅

---

No blockers. The Codex P1 finding is adequately addressed. Ready to merge.

[claude]

Claude finished @kyle-sexton's task in 59s —— View job

---

Security Review — PR #861

• Gather context and load tools
• Fetch PR diff
• Review changed files for security issues
• Post findings

---

No security issues found.

All four changed files are documentation and configuration metadata. Here is what was examined:

reference/pipeline-shape.md (+84/-0 / +12 in HEAD commit)
A new normative reference document. Contains no shell expansions, no template variable substitutions, no command execution, and no external URLs. The #NNN issue references are inert fragment anchors. No injection surface, no credential exposure, no PII.

skills/work/SKILL.md (+1/-1)
One sentence inserted in Step 5 pointing ${CLAUDE_PLUGIN_ROOT}/reference/pipeline-shape.md as the lane-shape authority — see permalink. ${CLAUDE_PLUGIN_ROOT} is the plugin's standard read-only path variable, already used throughout the file and system-controlled (not user-supplied). No path traversal risk, no new expansion or execution path, no new trust delegation.

plugin.json (+1/-1)
Semver version bump 0.18.2 → 0.19.0 only. No new permissions, no new userConfig keys, no new MCP server references, no egress changes, no supply-chain modifications.

CHANGELOG.md (+26/-0)
Changelog entries only. No security relevance.

No code execution paths, secrets handling, permission changes, dangerous GitHub Actions patterns, injection surfaces, or supply-chain risk introduced by this PR.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 91c0f89a81

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[claude]

Claude finished @kyle-sexton's task in 1m 45s —— View job

---

PR Review: feat(work-items): add mini-SDLC pipeline-shape SSOT + wire work Step 5 to it

• Gather context and classify request
• Read changed files at HEAD (ea33ac2)
• Review merge commit — confirm no conflicts or regressions introduced
• Review pipeline-shape.md at current HEAD
• Review work/SKILL.md at current HEAD
• Review plugin.json and CHANGELOG.md
• Post findings

---

Summary

Three commits on this branch (6886035, 91c0f89, ea33ac2). The first two were reviewed comprehensively in prior passes — no new concerns there. This review focuses on the merge-from-main commit (ea33ac2) and confirms the full PR is clean at the current HEAD.

---

Merge commit — ea33ac2: Merge origin/main into feat/613-mini-sdlc-pipeline-ssot

The merge pulled in 19 commits from main. The only conflict surface was the work-items plugin version — the Jira adapter (#379) had landed on main as 0.19.0 while this branch was targeting the same version number.

Version reconciliation:

File	main (0.19.0)	Branch HEAD (0.20.0)	Verdict
plugin.json	"version": "0.19.0"	"version": "0.20.0"	✅ Correct MINOR bump
CHANGELOG.md	## [0.19.0] (Jira adapter) at top	## [0.20.0] (pipeline-shape) above ## [0.19.0]	✅ Correct ordering

The bump is semantically correct: 0.20.0 is a MINOR release (new feature: new reference file + SKILL.md pointer) sitting above the Jira adapter's 0.19.0. No changelog entry is missing or misplaced. ✅

Observation — PR description version references are now stale. The description reads "0.18.2 → 0.19.0" throughout, reflecting the pre-merge intent. The actual bump landed as 0.19.0 → 0.20.0. This is cosmetic only — the committed files are correct — but a reader comparing the description to plugin.json will see a discrepancy. Not a merge blocker.

---

pipeline-shape.md at HEAD

ea33ac2 — 84 lines, unchanged from 91c0f89.

All findings from prior passes carry forward:

• Target-state voice maintained throughout. ✅
• Standalone-install fallback added in 91c0f89 addresses the Codex P1 finding. ✅
• STOPGAP status, OPERATOR DECISION markers, and placeholder tier names are prominent. ✅
• ${CLAUDE_PLUGIN_ROOT} path reference in work/SKILL.md is correct. ✅
• Plan-lane approval-gate tension noted in Codex P2 is correctly deferred to #513/#715 (dispatcher stage). ✅

---

work/SKILL.md at HEAD

ea33ac2 — one sentence inserted in Step 5, purely additive. No merge conflict; the sentence sits cleanly alongside the existing instruction chain. ✅

---

CLAUDE.md design-rule alignment

Rule	Status
Repo-agnostic (no hardcoded paths)	✅
${CLAUDE_PLUGIN_ROOT} for within-plugin refs	✅
Configurable without editing the plugin	✅
Plugin-form-safe (no ../ reach-outs)	✅
No PII / secrets	✅
Versioned with explicit semver	✅ (0.20.0)
Security (no code execution, no egress, no remote MCP)	✅

---

No blockers. The merge commit cleanly reconciled the version collision with the Jira adapter. The one observation (stale version reference in the PR description) is cosmetic and does not affect the committed files. Ready to merge.

[claude]

Claude finished @kyle-sexton's task in 1m 14s —— View job

---

Security Review — PR #861

• Gather context and classify request
• Fetch PR diff
• Review changed files for security issues
• Post findings

---

No security issues found.

All four changed files are documentation and policy content. Here is what was checked:

plugin.json (+1/-1) — version bump 0.19.0 → 0.20.0 only. No new permissions, no new userConfig keys, no new MCP server references, no egress changes. Existing userConfig entries (work_dispatch_concurrency_cap, work_cycle_batch_cap) are unchanged.

reference/pipeline-shape.md (+84/-0) — new normative reference document defining the per-item pipeline lane catalog and role-separation invariant (permalink). Contains no shell expansions, no template variable substitutions, no command execution, and no external URLs. The #NNN issue references are inert GitHub fragment anchors. No injection surface, no credential exposure, no PII. The file will be read as instruction content by the Claude Code runtime — this is the intended use case; the content is normative policy text with no user-controlled interpolation or adversarial prompt patterns.

skills/work/SKILL.md (+1/-1) — one sentence added to Step 5 pointing ${CLAUDE_PLUGIN_ROOT}/reference/pipeline-shape.md as the lane-shape authority (permalink). ${CLAUDE_PLUGIN_ROOT} is a system-controlled read-only path variable already used throughout this file — not user-supplied and not externally injectable. The path is a static within-plugin reference with no ../ traversal. No new trust delegation, no new execution path.

CHANGELOG.md (+26/-0) — changelog entries only. No security relevance.

No code execution paths, secrets handling, permission changes, dangerous GitHub Actions patterns, injection surfaces, or supply-chain risk introduced by this PR.

[kyle-sexton]

This was generated by AI (decide lane — decision-brief pass; parked PR-thread decision surfaced via periodic scan).

Decision brief — Plan lane vs. autonomous execution. Parking a decision surfaced in review thread discussion_r3624454271 (Codex P1 "Preserve autonomous execution when plan approval is unavailable"). The author verified the tension as VALID and left the thread open, deferring to the #513/#715 dispatcher stage without marking a remedy RECOMMENDED. This brief rules the constraint now and scopes the mechanism to #715.

Problem

The shape's mandatory Plan lane routes every autonomous /work-items:work item through /planning:plan, whose Step 5 is a human-approval gate. If the future dispatcher runs that lane at full realization, an unattended run parks the moment it finishes planning and never reaches implementation — silently, after already claiming the item and holding a lease.

Today

• Plan lane = /planning:plan. pipeline-shape.md:41 names it the lane's realization; pipeline-shape.md:24 — "A lane is never skipped," so every autonomous item hits this lane.
• The gate is human-specific and terminal. planning/skills/plan/SKILL.md:210 "the user approves, modifies, or rejects it before execution begins"; :206 "the user is the terminal gate (deterministic check → human)"; :341 "the approval gate is the point — it's where human judgment enters the loop." The gate's designed purpose is injecting human judgment, not merely a second pair of eyes.
• /work-items:work claims then dispatches autonomously. work/skills/work/SKILL.md:181 — orchestrator claims + dispatches after acquiring the lease; there is no approval handoff or autonomous planning path. The autonomous frontier filter list-frontier --autonomous (:130) only excludes the needs-human role label — it does not mean "this plan is pre-approved."
• Nothing reads the shape yet. The dispatcher that would route the lane is #715 (Stage 2), unbuilt (pipeline-shape.md:74); pipeline-shape.md makes no claim about runtime today. So the conflict is latent, not live.

Blocked or ruleable?

Split. The constraint is ruleable now; the mechanism is blocked on #715 — and blocked by a real design dependency, not just ownership. pipeline-shape.md:25-27 says a lane's realization collapses (fresh-context subagent → inline lightweight check → consumer's own workflow step). A shallow-tier Plan lane may be a gateless inline check that never invokes /planning:plan Step 5 at all. So whether the gate even fires depends on how #715 specifies the shallow-Plan collapse. Codex's "parks immediately" is therefore conditional — it assumes the collapsed Plan lane still routes through full /planning:plan. The mechanism can't be correctly chosen until #715 settles that collapse spec.

Options

➡️ RECOMMENDED — Record the binding constraint now; defer the mechanism to #715/#572.
Rule as policy, in #715's frame: an autonomous run must never silently park at a human-judgment gate, and whichever mechanism #715 picks must preserve independent human judgment over the plan (producer ≠ approver — the plan's author cannot self-approve). This is the tie-breaker across the three remedies below and subsumes the author's "defer" disposition while making it actionable — #715 inherits a constraint, not an open question. Reversibility: high (a recorded constraint, no runtime committed). If autonomous dispatch (#572) ships before #715, adopt Option 2 as the interim posture in the meantime.

Option 1 — Autonomous-safe planning variant (gateless Plan lane). Changes the skill the lane uses. Rejected as a standalone. planning/skills/plan/SKILL.md:341/:206 make the gate specifically a human-judgment injection; a gateless variant doesn't just let the producer self-approve — it deletes the human judgment the gate exists for, violating the invariant this whole PR holds (implementer ≠ reviewer ≠ verifier, extended to producer ≠ approver). Viable only if it substitutes an independent approver — which is Option 3, not a bare gate removal. Reversibility: medium (new skill surface to unwind).

Option 2 — Stop-before-claim (interim posture). Don't claim work that would stall at the gate. Caveat, not a free reuse: since the Plan lane is never skipped (:24), "skip items needing planning" would starve the lane of everything. It only works as a policy extension — move the human's approval to triage-time eligibility (an item is marked autonomous-eligible only once its plan/approach is pre-approved), pre-satisfying the plan gate for eligible items. Today's needs-human is a role label, not "this plan needs approval," so this extends the triage state machine, it does not merely reuse the existing filter. Reversibility: high (label/selection policy, no new mechanism). Best interim answer if dispatch ships before #715.

Option 3 — Approval-handoff mechanism. The autonomous run parks the plan to an async approval surface (human, or a designated distinct approver) and resumes on approval — never discarding the work. Preserves producer ≠ approver and keeps the work alive; it is the richest durable answer, and the natural home for #715/#572 to build. Reversibility: low (a new mechanism + lifecycle state). Do not build ahead of #715's collapse spec.

Cross-links

• #513 / #715 — mini-SDLC umbrella / Stage-2 Opus complexity-dispatcher; owns lane routing and the shallow-Plan collapse spec this decision depends on.
• #572 — autonomous orchestrator-dispatch lifecycle (branch/worktree provisioning, PR-creation ownership); the approval-handoff / stop-before-claim seam lives on this same autonomous lifecycle surface.
• #304 — fresh-eyes / producer ≠ reviewer checkpoint; the same-context-bias principle that grounds "producer ≠ approver" here.

[kyle-sexton]

Tower merge (delegation, veto window open): CLEAN, all checks green, 0 unresolved threads (all rounds classified + resolved), version restacked to 0.20.0 post-collision. Closes #613 — mini-SDLC stage 1 SSOT lands. This was generated by AI (control tower, session 6).