Skip to content

Latest commit

 

History

History
539 lines (472 loc) · 60.9 KB

roadmap.md

File metadata and controls

539 lines (472 loc) · 60.9 KB

Roadmap

Two parts. The product roadmap below is what to build or fix next for Perun, produced by the CEO improvement loop: ranked by a printed formula from sourced-or-assumed inputs, with a success metric, evidence links, ADRs and tracker-ready issues. The suite design record after the divider is the 2026-09-11 design of the agent suite, kept unchanged.

To regenerate, edit roadmap-input.json and run python3 .claude/skills/agentic-ceo/scripts/roadmap_rank.py docs/roadmap-input.json --update docs/roadmap.md. scripts/test_roadmap_rank.py fails when this page and its input drift apart.

BLUF. Stage mvp (assumed). 19 items ranked (11 open, 8 in flight), 9 rejected. Top: sandbox-preflight 3.75 (in flight), gate-cache-tree-sha 3.75 (in flight), pr-state-one-shot 1.88 (in flight). Start now (highest-ranked open item with no gate): background-work-guard. 11 of 19 reach inputs are assumed; equal scores rank fewer assumed inputs first.

1. Strategy context

As of 2026-10-16. A source that starts with assumed is an explicit assumption, not a repo fact.

Field Value Source
stage mvp assumed: no repo doc states a stage. Signals: version 1.586.0, a public MIT repo, CI, gates and evals in place, a small set of peer operators (survey:2026-10-16), no retention or product-market-fit data and unverified adoption (docs/perun-vs-rivals.md section 3), five skills ahead of their rivals on preregistered fresh held-out sets with LLM judges on synthetic scenarios from one generator family (agentic-ceo, positioning and product-discovery replicated; agentic-delivery and product-output-safety narrow; a one-run idea-critic win did not replicate), the others at parity or behind (docs/perun-vs-rivals.md, top table). A growth stage needs a product-market-fit read (product-discovery); none exists.
vision One solution across code review, token optimisation, delivery, product work and C-suite advice, measurably better than rival bundles in each area, with every claim kept inside the evidence. owner:2026-10-16 (stated goal). README.md states the review-first version and docs/roadmap.md the suite version (honesty, verification, stage-awareness).
north star Tokens per delivered, gate-green PR, with review quality (F1 on the public case set) held constant. assumed: no repo doc names a north star. Closest instruments: .claude/skills/agentic-ceo/scripts/token_ratchet.py (tokens per delivered PR) and the review F1 in docs/perun-vs-rivals.md; neither is measured across installs, because there is no telemetry.
direction Close the measured gaps against rivals (review cost at equal quality; the product skills still at parity or behind their rivals), remove the delivery friction peers report, and keep every claim inside the evidence. docs/roadmap.md (suite direction and epistemic spine), docs/perun-vs-rivals.md (measured gaps), survey:2026-10-16 (peer pain points)
constraints No paid model benches in this pass. Instructions, not a service: no telemetry. No fabricated evidence; the privacy and no-duplication rules apply. In-flight work is listed, not re-ranked. owner:2026-10-16 (no paid benches); README.md (Safety: instructions, not a service); CLAUDE.md (hard rules)

Owner question (one batch): confirm or correct the assumed stage, north star. The run continues on the assumption; a correction re-runs this script.

Evidence base (a source that was absent or unreadable is listed as such, never skipped):

  • Run mode: CEO direct. One agent applied the CPO, CTO, CFO, CMO, CISO and COO briefs in turn to the same evidence and spawned no lanes, because the evidence fits one context and a fan-out would only add cost. The CISO lens found no open item with a security lever that clears the evidence bar; its host-safety concerns are folded into janitor-teardown-land and handback-automator.
  • Peer survey, 2026-10-16, rows P1 to P10, and field measurements F1 to F3, committed as docs/peer-survey-2026-10-16.md: an operator-supplied relay, not a measurement made in this repo, without per-item counts. P1 hook reruns of 2 to 15 minutes; P2 gates run 3 times per tree; P3 the sandbox blocks gh, Chromium, Postgres and tsx IPC; P4 lanes end their turn while background tests still run; P5 flaky tests drop PRs from trains; P6 a pr_state one-shot command is wanted; P7 a lane-handback automator is wanted; P8 a sandbox preflight is wanted; P9 reply-cap rewrites; P10 re-deriving state after compaction. F1 a 47-minute hook cycle fell to 30 seconds with gate dedup; F2 about 5 sandbox retries per session; F3 zero-chat lanes, no measurement. A survey row or relayed number counts as a source for an input, but reach stays a hypothesis until counts exist.
  • GitHub issues and PRs for protocol/perun, REST API, fetched 2026-10-16: 14 open issues (61, 69, 267, 273, 393, 1316, 1380, 1418, 1430 to 1433, 1530, 1531) and 1 open PR (1382, last updated 2026-10-07). No open PR exists for any in-flight item the owner named, so those statuses are owner-stated and unverified on GitHub.
  • CHANGELOG.md, releases 1.584.0 to 1.586.0 (merge-train flake rules, extra hosts, auto-compact with handoff re-inject): read to reject work that already shipped.
  • docs/perun-vs-rivals.md: the repo's own benchmarks H1 to H11 and the rival gaps in section 3. Review ties the Anthropic toolkit at about 3.4 times the cost (not harness-matched: the toolkit ran as --system-prompt; the only matched pair, 1.05 times, is the unshipped V2 prompt, so the shipped skill is unmeasured; docs/bench/review-cost-matched-1023/PREREGISTERED.md), and five skills beat their rivals on preregistered fresh held-out sets with LLM judges on synthetic scenarios from one generator family (agentic-ceo, positioning and product-discovery replicated; agentic-delivery and product-output-safety narrow; a one-run idea-critic win did not replicate) while the others are at parity or behind.
  • Rival and platform changelogs, fetched live 2026-10-16: obra/superpowers releases v7.0.0, v6.4.2, v6.4.1 and v6.3.0 (GitHub releases API) and the Claude Code CHANGELOG.md, top entry 2.1.301 (raw.githubusercontent.com). UNVERIFIED: gstack releases and the Claude Code release pages, because the unauthenticated GitHub API rate limit was hit; not retried.
  • Not present: analytics (Perun is instructions, not a service and has no telemetry), support threads and a user-feedback file. Not read: session logs (private by default); token_report.py and weekly_receipt.py are the instruments for an operator's own repo. docs/capability-matrix.md does not exist at origin/main c097d01f (checked 2026-10-16), so it was not used.

2. Rejected items

Item Rule Reason
flake-ledger-quarantine Per-test flake ledger with quarantine, so flaky tests stop dropping PRs from trains CEO decision Already shipped: per-test flake ledger with quarantine (CHANGELOG.md 1.587.0, flake_ledger.py, train_flake.sh and train_land.sh read the quarantine list; scripts/test_flake_ledger.py proves a consistent failure is never quarantined).
usage-telemetry Opt-out usage telemetry to measure reach vision: misaligned Breaks a stated constraint: Perun is instructions, not a service (README.md, Safety). The cheaper route to measured reach is a survey with counts (peer-survey-v2) and the operator-local receipts that already exist (weekly_receipt.py).
rederive-state-after-compaction Re-derive state after compaction (survey P10) as a new feature CEO decision Already shipped: auto-compact with handoff re-inject (CHANGELOG.md 1.586.0, handoff_inject.py) and the task ledger (task_ledger.py). The remaining gap is PR state, covered by the in-flight pr-state-one-shot.
compact-card-recall-variants More compact-card variants to recover review recall CEO decision A measured dead end: the best of three variants failed on the held-out half (n=22, F1 -0.036, interval -0.194 to +0.095, H7), and the real skill already ties the toolkit on recall (H7b). Reopen only with a new idea and a held-out plan.
own-output-safety-eval-runner Build a product-output-safety eval runner inside Perun CEO decision Better way, registry not a bundle: an existing tool already runs evals and red-team scans in CI (promptfoo, section 3). Ship a config template only if the H1b loss is shown to come from missing tooling; that test comes first (differentiator-bench).
host-count-expansion Add more native hosts CEO decision Breadth before depth: 18 of the 19 listed hosts were never run on a real host (H10) and the one open candidate, Muse, had no vendor docs (CHANGELOG.md 1.585.0). At mvp the stage weights favour verifying what is listed (real-host-smoke-tests).
session-diagnosis-skill Add a session-diagnosis skill, as a rival did CEO decision Deferred on evidence, not rejected on merit: one rival release is a lead, not user evidence (product-discovery riskiest-assumption gate), and token_report.py covers spend, not causes. One question goes into peer-survey-v2 (how is a failed session diagnosed today?); reopen on the answers.
autonomy-advisor Autonomy advisor: read the sandbox and permission settings and print the narrow rules that stop sandbox retries, never writing settings CEO decision Already shipped: sandbox_preflight.py prints the narrow rules (RISKS keys tsx, browser, db, gh) and sandbox_probe.py now probes gh, headless Chromium, a Postgres client connect and tsx at runtime, naming the preflight rule that fixes each failure (SKIP when a tool is absent, exit 1 on a real block, nothing installed or written).
zero-chat-lanes Zero-chat lanes: a lane's chat is one status line, everything else goes to files CEO decision Already shipped: fields-only hand-backs, the injected lane rule and the hand-back audit landed in 1.587.0 (scripts/test_zero_chat_lanes.py).

3. Ranked roadmap

Score = reach × impact × confidence ÷ effort × stage weight × vision factor. The score orders the list, it does not prove anything. The loop scores itself, so every input that is not backed by a real link or repo path is listed as assumed below.

  • reach 1-5: share of our own users or runs touched, 5 = every one (never a market size; an item that only helps later rankings has no reach)
  • impact 0.25, 0.5, 1, 2 or 3: minimal, low, medium, high or massive effect on the north star
  • confidence is computed: 1 when reach and impact both cite a URL, an existing repo path or an issue or PR, else 0.5; assumed, estimate: or any other text counts as assumed
  • effort: lane-days, plus 2 per owner or external-send gate and 0.25 per calendar-wait day
  • stage weight: lever × stage from the table below; the active stage is mvp
  • vision factor: aligned 1.0, partial 0.5; a misaligned item is rejected in section 2
  • items already in flight are scored and flagged, not dropped; equal scores rank fewer assumed inputs first, then lower effort, then id
lever prototype mvp (active) growth mature
learning 1.5 1.5 1 0.75
reliability 1 1.25 1.5 1.5
velocity 1.25 1.25 1 0.75
growth 0.25 0.5 1.5 0.75
scale 0.25 0.5 1.25 1.5
cost 0.5 0.75 1.25 1.5
polish 0.25 0.5 1 1.25
security 1.5 1.5 1.5 1.5
# Item Status Lane, lever Score Success metric Evidence Gate
1 sandbox-preflight Sandbox preflight for gh, Chromium, Postgres and tsx IPC, with the reply-cap code-block exemption in flight -, reliability 3.75 Blocked-tool surprises per lane from 4 named tools (survey P3) to 0 unreported: one verdict line per tool before work starts; reply-cap rewrites (survey P9) counted before and after the code-block exemption lands. survey:2026-10-16#P3; survey:2026-10-16#P8; survey:2026-10-16#P9; sandbox_probe.py; gh_rest.py; https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md -
2 gate-cache-tree-sha Shared tree-SHA gate cache with gate_map and light-lane weighting in flight -, velocity 3.75 Hook cycle 47 minutes to 30 seconds and gate runs per tree 3 to 1 (field report F1; survey P1, P2); report the numbers when it lands. survey:2026-10-16#P1; survey:2026-10-16#P2; pre-push-verify.sh; preflight.py -
3 pr-state-one-shot pr_state: one command for a PR's checks, mergeability and review state in flight -, velocity 1.88 Commands needed to learn a PR's state from several gh calls (survey P6) to 1; state after a compaction read in 1 call (survey P10). survey:2026-10-16#P6; survey:2026-10-16#P10; gh_rest.py -
4 handback-automator Lane-handback automator: gates, required-check names, base-branch fast-forward and tracker post; cleanup printed, never run in flight COO, velocity 1.25 Manual steps per handback: 6 (lane-preamble.md: wait, handback, clean_finished.sh, reap_own.sh, changelog fragment, git diff --check) to 1; handbacks needing a second pass: unmeasured, count for 2 weeks. survey:2026-10-16#P7; lane-preamble.md; issue:#1530; issue:#1531 -
5 low-context-skill-index Low-context skill index in flight -, cost 1.25 Always-loaded tokens, full install 921 and review-only install 75 (section 2), and review cost per case $0.0551 (H7b): report all three after it lands; the routing eval stays at 24 of 30 (H3). perun-vs-rivals.md#2-always-loaded-toke…; perun-vs-rivals.md#h7b-review-patch-onl…; INDEX.md -
6 background-work-guard Block a handback while any background test or build from the worktree is still running open COO, reliability 0.94 Lanes handing back with a live background process: unmeasured now (survey P4), target 0; the guard blocks the planted fixture 1 of 1. survey:2026-10-16#P4; handback_cap.py; lane-preamble.md; issue:#1530 -
7 janitor-teardown-land Land the worktree janitor (PR 1382, open since 2026-10-07) and the teardown standard (issue 1531) open COO, reliability 0.94 Finished worktrees and leftover dev servers at session start: 130 or more in the field report behind session_brief.py, target 0 after a train; PR 1382 age 9 days to merged or closed. pr:#1382; issue:#1531; session_brief.py; https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md -
8 differentiator-bench Benchmark the claimed differentiator (refusing to fabricate) with deterministic predicates, not a rubric judge open CPO, learning 0.75 Bait-prompt pass rate, Perun vs the best rival, graded without a model: unmeasured now; n of 20 or more with a 95% interval; claim superiority only if the interval excludes 0. roadmap.md; self-improvement.md; eval_predicates.py; perun-vs-rivals.md#h1d-idea-critic-prod… owner: a small paid run under a cap, and a rival skill installed
9 review-cost-per-recall Cut review cost per unit of recall: ratchet the review path's tokens, then re-bench once open CFO, cost 0.62 Review cost per case at equal F1 vs the Anthropic toolkit: 3.4x ($0.0551 vs $0.0162, n=22, H7b, not harness-matched: toolkit ran as --system-prompt, re-measure with docs/bench/review-cost-matched-1023) to 2x or less; web-archetype must-load ceiling 20,049 estimated tokens, ratcheted down. perun-vs-rivals.md#h7b-review-patch-onl…; README.md; mustload-budgets.tsv; https://github.com/obra/superpowers/releases/tag/v6.4.1 owner: the paid re-bench
10 host-command-hints Host-aware command, model and effort hints in flight -, reliability 0.62 Hosts with a hint verified on a real host: 1 of 19 at the start (H10); report the new count when it lands. owner:2026-10-16; perun-vs-rivals.md#h10-install-and-upda… -
11 derived-numbers-gate Extend the numbers-need-a-source gate from the leaders page to the README and the rivals page open CTO, reliability 0.62 Pages under the digit-needs-a-source rule: 1 of 3 now (docs/for-leaders.md only), 3 of 3. check-leader-doc.sh; perun-vs-rivals.md; CHANGELOG.md -
12 review-lenses Review lenses (scope as stated by the owner) in flight -, reliability 0.62 Open lens issues 1430 to 1433: 4 open today; each closed with a planted-defect eval. owner:2026-10-16; issue:#1430; issue:#1431; issue:#1432; issue:#1433 -
13 loop-before-after-measure Measure the improvement loop against a plain 'what should we build next' prompt on two real repos open CPO, learning 0.60 Win rate vs a plain prompt on a rubric-free judge: unmeasured now; 20 or more pairs; claim a win only if the lower bound is above 0.5. improvement-loop.md; perun-vs-rivals.md#h3-skill-routing-age…; bench owner: paid bench
14 judge-independence-rerun Re-run the rubric-free comparisons with a judge from another vendor family open CPO, learning 0.38 Re-tested pairs judged by a different-vendor judge: 0 of 5 now, 5 of 5; per pair, whether the outcome moved by more than its interval. perun-vs-rivals.md#what-was-not-measured; perun-vs-rivals.md#h1d-idea-critic-prod… owner: paid calls and a second vendor's key
15 url-switch URL switch to protocol/perun and the plugin rename in flight -, polish 0.31 Lines naming the previous repository URL in tracked files outside CHANGELOG.md: 23 in 14 files (git grep at c097d01f) to 0 outside fixtures and tests that name it on purpose; plugin name still deep-code-review in .claude-plugin/plugin.json. README.md; plugin.json; perun_auto_update.py; owner:2026-10-16 -
16 distribution-listing Distribution: an official marketplace listing and positioning against competing prose skills (issue 273) open CMO, growth 0.27 Verified marketplace listings: 0 of 1 now (status unverified, section 3), 1 of 1; no usage target, since installs are not countable. issue:#273; perun-vs-rivals.md#3-what-the-rivals-ha…; perun-vs-rivals.md#h10-install-and-upda… owner: a marketplace submission is an external send
17 real-host-smoke-tests Smoke-test more of the listed hosts on real hosts, starting with the ones that cost nothing to run open CTO, reliability 0.25 Hosts smoke-tested on a real host (11 of 11 skills listed): 1 of 19 now, 5 of 19. perun-vs-rivals.md#h10-install-and-upda…; host-safety.md; CHANGELOG.md owner: vendor accounts or CLIs for each host
18 peer-survey-v2 Peer survey v2: per-item counts, a very-disappointed question and a week-4 still-using question open CPO, learning 0.23 Open items with a sourced (not assumed) reach: 0 now, at least half at the next run; product-market-fit read: UNVERIFIED now, a very-disappointed share with n stated. survey:2026-10-16; SKILL.md; perun-vs-rivals.md#3-what-the-rivals-ha… owner: sending the form is an external send
19 product-skill-runnable-checks Give each product skill one runnable check or template, as the rivals ship open CPO, growth 0.05 Product skills with a runnable script: 0 of 5 now (product-discovery, growth-analytics, positioning, business-ops, product-output-safety); 2 of 5 only if the differentiator bench shows a tooling gap. perun-vs-rivals.md#3-what-the-rivals-ha…; perun-vs-rivals.md#h1d-idea-critic-prod… -

Score inputs:

# reach impact confidence effort (typed + gates + wait) stage × vision × score assumed (R reach, I impact, E effort)
1 3 2 1 2 1.25 1 3.75 E
2 3 3 1 3 1.25 1 3.75 E
3 3 1 0.5 1 1.25 1 1.88 I E
4 3 2 0.5 3 1.25 1 1.25 I E
5 5 2 0.5 3 0.75 1 1.25 I E
6 3 1 0.5 2 1.25 1 0.94 I E
7 3 1 0.5 2 1.25 1 0.94 I E
8 2 2 0.5 2 + 2 + 0 = 4 1.5 1 0.75 R I E
9 5 2 0.5 4 + 2 + 0 = 6 0.75 1 0.62 R E
10 2 1 0.5 2 1.25 1 0.62 R I E
11 3 1 0.5 3 1.25 1 0.62 R I E
12 3 1 0.5 3 1.25 1 0.62 R I E
13 2 2 0.5 3 + 2 + 0 = 5 1.5 1 0.60 R I E
14 2 1 0.5 2 + 2 + 0 = 4 1.5 1 0.38 R I E
15 5 1 0.5 2 0.5 0.5 0.31 I E
16 5 3 0.5 5 + 2 + 0 = 7 0.5 0.5 0.27 R I E
17 2 1 0.5 3 + 2 + 0 = 5 1.25 1 0.25 R I E
18 1 2 0.5 1 + 2 + 3.5 = 6.5 1.5 1 0.23 R I E
19 2 1 0.5 5 0.5 0.5 0.05 R I E

Top 3 open items attacked by idea-critic

    1. background-work-guard: PASS_TO_USER (inline). Strongest attack survived: A guard that blocks on any cwd process will hold lanes behind long-lived dev servers the user owns. It survives because the existing wait bound releases the hold, the check fails open when it cannot see, and the kill criterion drops it to advisory after 1 false block per 20. Kill criteria: More than 1 false block per 20 handbacks in the first 2 weeks: ship advisory only; ps or lsof unavailable on a host: fail open with COULD_NOT_CHECK, never block Remaining risk: A second class of processes the cwd check cannot see.
    1. janitor-teardown-land: PASS_TO_USER (inline). Strongest attack survived: A cleanup tool that deletes is itself a host-safety risk, and a past incident deleted user files. It survives only as report-first, acting solely on paths the run recorded creating, with exact-PID teardown and no variable-path recursive delete. Kill criteria: The janitor ever proposes a path it did not create: stop and keep report-only; Report-only shows fewer than 5 leftovers per train over 3 trains: close PR 1382 as not worth the risk Remaining risk: PR 1382 may be stale against main, and leftover counts come from one field report.
    1. differentiator-bench: PASS_TO_USER (inline). Strongest attack survived: The predicates are heuristics tuned on Perun's own fixtures, so they could favour Perun, the same bias class as the rubric. It survives only because done-when requires calibration on rival outputs and a pre-registered decision rule before scoring; without that step the verdict would be REVISE. Kill criteria: The predicates fail to separate the rival golden pairs: stop, the instrument is not valid for rivals; Perun and the best rival pass at the same rate (interval contains 0): drop honesty as the stated differentiator in positioning and do not widen the claim Remaining risk: The predicates stay narrower than real fabrication; a tie is a valid and useful result.

Dedup check: 28 items, near-duplicate titles (similarity 0.85 or more): none.

4. Technical-direction decisions (ADRs)

ADR-2: Compose the existing handback scripts in one thin entry point (tooling), serves handback-automator

  • Status: proposed
  • Context: A handback is six manual steps today (lane-preamble.md): wait, handback, clean_finished.sh, reap_own.sh, the changelog fragment and git diff --check. Each step already has a script and tests.
  • Options: (A) A thin script that calls the existing scripts in order and prints the cleanup: No duplicated logic; one more entry point to keep in step with lane-preamble.md. (B) Grow lane_guard.py handback: One entry point; mixes the start-of-lane guard and the end-of-lane automation in one file. (C) Do nothing, prose only: No code; the survey (P7) still asks for automation.
  • Decision: We will add a thin composer that calls the existing scripts and prints, never runs, the worktree cleanup and any base-branch fast-forward not covered by a human push grant, because every step already has a tested script and shared-state actions keep their human gate.
  • Consequences: Good: fewer steps with no new gate logic. Bad: a second entry point to maintain. Neutral: dry run is the default, so the first release automates less than the survey asked for.
  • Reversal cost: Low: delete one file and its preamble line.

ADR-3: Enforce the background-work rule at the SubagentStop hook (architecture), serves background-work-guard

  • Status: proposed
  • Context: lane-preamble.md says never end a turn with background jobs running, and the survey (P4) says lanes still do. handback_cap.py already blocks a handback while a git push from the worktree is live, by matching PUSH_RE against the process list and checking the cwd with lsof, but it matches nothing else.
  • Options: (A) Extend handback_cap.py's process check to any process with its cwd in the worktree: Reuses a tested mechanism with its bounded wait and fail-open path; may hold a handback on a long-lived dev server. (B) Keep the rule in prose: No code; it has not held (survey P4). (C) Register every background launch in a wrapper: Exact, but needs a harness feature and touches every lane.
  • Decision: We will extend the existing check to any process whose cwd is inside the worktree, with the same wait bound and the same COULD_NOT_CHECK fail-open path, because the hook already runs at the right moment and holds the cwd logic.
  • Consequences: Good: closes the gap where the push check is already enforced. Bad: a user-owned dev server inside the worktree can hold a handback until the bound expires. Neutral: the nested-worktree bug in issue 1530 must be fixed first or together.
  • Reversal cost: Low: revert one function; no data or format is involved.

ADR-4: Store survey counts as an aggregate CSV in the repo (tooling), serves peer-survey-v2

  • Status: proposed
  • Context: roadmap_rank.py reads each input's source as a path, a link or the word assumed, so measured reach needs a stable file that carries no personal data.
  • Options: (A) Aggregate CSV under docs/, counts and n only: Versioned with the roadmap, reproducible and checkable by the privacy gate; no raw answers are kept. (B) Counts in an issue comment: Easy to post; not versioned with the roadmap and lost if the tracker changes. (C) A form tool's dashboard: Zero setup; third-party data that cannot be cited or re-derived from the repo.
  • Decision: We will commit an aggregate CSV of counts and n only under docs/ and cite it from reach_src, because that keeps the roadmap reproducible and carries no names.
  • Consequences: Good: measured reach is checkable in review. Bad: raw answers are not kept, so a question cannot be re-sliced later. Neutral: the form itself stays outside the repo.
  • Reversal cost: Low: the CSV is data; delete it and reach goes back to assumed.

5. Tracker-ready issues

Open items only (in-flight work already has its issue). Priority is rank-derived among open items (ranks 1-3 P1, 4-6 P2, the rest P3); P0 is a human call and is never assigned here. Issues are drafted for P1 and P2 only; a P3 item stays in the table until it is promoted.

P1: Block a handback while any background test or build from the worktree is still running

  • Labels: reliability, COO, stage:mvp
  • Where: .claude/skills/agentic-delivery/scripts/handback_cap.py (the SubagentStop hook; its process check matches git push only).
  • Done when: The hook blocks while any process whose cwd is inside the worktree is alive, not only git push, bounded by the existing wait limit; when ps or lsof cannot answer it still fails open with COULD_NOT_CHECK; the nested-worktree case from issue 1530 is covered.
  • Verify: python3 .claude/skills/agentic-delivery/scripts/handback_cap.py --selftest with a live sleep-process fixture and a nested-worktree fixture.
  • Why: Delivery: the rule exists in prose (lane-preamble.md) and the survey shows it does not hold; enforce it where the push check already lives. (vision: aligned)
  • Success metric: Lanes handing back with a live background process: unmeasured now (survey P4), target 0; the guard blocks the planted fixture 1 of 1.
  • Evidence: survey:2026-10-16#P4; handback_cap.py; lane-preamble.md; issue:#1530
  • Risks: False blocks on long-lived dev servers the user owns: scope the check to the worktree and keep the release valve.
  • Merged from: COO lens: survey P4; CTO lens: handback_cap.py matches git push only (PUSH_RE)
  • Score: 0.94 = 3 × 1 × 0.5 ÷ 2 × 1.25 × 1 (open rank 1 of 11); assumed: impact effort

P1: Land the worktree janitor (PR 1382, open since 2026-10-07) and the teardown standard (issue 1531)

  • Labels: reliability, COO, stage:mvp
  • Where: PR 1382 (janitor), the teardown standard in issue 1531, .claude/skills/agentic-delivery/scripts/clean_finished.sh and reap_own.sh.
  • Done when: The standard (exact-PID teardown, EXIT traps, atomic claims) is written and the janitor follows it: it reports and prints cleanup commands first and removes only what it proves it owns; PR 1382 is merged or closed.
  • Verify: A fixture with one owned and one foreign worktree plus a foreign dev server: the janitor removes or prints only the owned one and never signals the foreign process; session_brief.py reports 0 finished worktrees afterwards.
  • Why: Delivery: stale worktrees and orphan dev servers are the cleanup cost of every fleet run, and the platform now guards hook-made worktree removal too (Claude Code 2.1.301), so Perun's own handling should not be weaker. (vision: aligned)
  • Success metric: Finished worktrees and leftover dev servers at session start: 130 or more in the field report behind session_brief.py, target 0 after a train; PR 1382 age 9 days to merged or closed.
  • Evidence: pr:#1382; issue:#1531; session_brief.py; https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md
  • Risks: A janitor that deletes is itself a host-safety risk: report and print before it removes, and never act on a path it did not create.
  • Merged from: COO lens: stale worktrees (session_brief.py); CISO lens: a deleting janitor is a host-safety risk
  • Score: 0.94 = 3 × 1 × 0.5 ÷ 2 × 1.25 × 1 (open rank 2 of 11); assumed: impact effort

P1: Benchmark the claimed differentiator (refusing to fabricate) with deterministic predicates, not a rubric judge

  • Labels: learning, CPO, stage:mvp
  • Where: scripts/eval_predicates.py (bindings), the product skills' evals, docs/bench/ (a preregistration for the run).
  • Done when: 5 golden refusals and 5 golden fabrications taken from rival runs separate under the predicates; the decision rule is committed before scoring; then at least 20 bait prompts run through Perun's product skills, one rival skill and a plain prompt, graded by the predicates, with a 95% interval per pair.
  • Verify: python3 scripts/eval_predicates.py --selftest stays green; the preregistration commit precedes the run in git log; the report states each pair's interval and says tie when it contains 0.
  • Why: Product skills: tests the axis Perun says it wins on, the 'superior to rivals in each' goal, with a grader that cannot be rubric-biased. (vision: aligned)
  • Success metric: Bait-prompt pass rate, Perun vs the best rival, graded without a model: unmeasured now; n of 20 or more with a 95% interval; claim superiority only if the interval excludes 0.
  • Evidence: roadmap.md; self-improvement.md; eval_predicates.py; perun-vs-rivals.md#h1d-idea-critic-prod…
  • Risks: The predicates are heuristics tuned on Perun's own fixtures and could favour Perun; calibrate them on rival outputs first. A tie is a valid result and would retire honesty as the stated differentiator in positioning.
  • Gate: owner: a small paid run under a cap, and a rival skill installed
  • Merged from: CPO lens: no rubric-free win (H1c, H1d); CMO lens: the stated differentiator is honesty (docs/roadmap.md)
  • Score: 0.75 = 2 × 2 × 0.5 ÷ 4 × 1.5 × 1 (open rank 3 of 11); assumed: reach impact effort

P2: Cut review cost per unit of recall: ratchet the review path's tokens, then re-bench once

  • Labels: cost, CFO, stage:mvp
  • Where: scripts/mustload-budgets.tsv and the deep-code-review load map; the re-bench follows the H7b method.
  • Done when: The review path's must-load tokens have a ratchet row per archetype that only moves down; the low-context index has landed; one re-bench on the H7b case set reports cost and F1 with intervals.
  • Verify: bash scripts/ci-gates.sh mustload . stays green and the ceilings fall; the re-bench ledger is committed in docs/bench/ before the page is edited.
  • Why: Token optimisation and code review: review ties the cheapest rival on recall at about 3.4 times the cost, so cost is the measured gap to close. On the 37-case run (every arm at recall 0.297) a review costs about $0.46 per caught bug against $0.30 to $0.33 for the plain prompt and rivals (derived: review cost per case divided by recall; aggregates only, docs/review-cost-per-recall.md), and a rival's v6.4.1 added a cheaper execution mode (Native execution). (vision: aligned)
  • Success metric: Review cost per case at equal F1 vs the Anthropic toolkit: 3.4x ($0.0551 vs $0.0162, n=22, H7b, not harness-matched: toolkit ran as --system-prompt, re-measure with docs/bench/review-cost-matched-1023) to 2x or less; web-archetype must-load ceiling 20,049 estimated tokens, ratcheted down.
  • Evidence: perun-vs-rivals.md#h7b-review-patch-onl…; README.md; mustload-budgets.tsv; https://github.com/obra/superpowers/releases/tag/v6.4.1
  • Risks: Cheaper review that loses recall is not a win: F1 must hold within the H7b interval. The compact-card route failed on held-out cases (H7), so cut loaded tokens, not instructions.
  • Gate: owner: the paid re-bench
  • Merged from: CFO lens: the H7b cost gap; CMO lens: README cost per case against a plain prompt
  • Score: 0.62 = 5 × 2 × 0.5 ÷ 6 × 0.75 × 1 (open rank 4 of 11); assumed: reach effort

P2: Extend the numbers-need-a-source gate from the leaders page to the README and the rivals page

  • Labels: reliability, CTO, stage:mvp
  • Where: scripts/check-leader-doc.sh, README.md, docs/perun-vs-rivals.md.
  • Done when: The gate runs on all three pages; every body line with a digit links to a repo file or says unmeasured, or is on a short reviewed allowlist; CI fails on a planted bad line.
  • Verify: bash scripts/check-leader-doc.sh --selftest passes, the gate passes on the two new pages, and a planted bad line fails.
  • Why: Product and review claims: honesty is the stated differentiator, the rivals page was reworded in 1.586.0 to correct what it said about unpreserved rubrics, and the source-link rule covers only docs/for-leaders.md. (vision: aligned)
  • Success metric: Pages under the digit-needs-a-source rule: 1 of 3 now (docs/for-leaders.md only), 3 of 3.
  • Evidence: check-leader-doc.sh; perun-vs-rivals.md; CHANGELOG.md
  • Risks: The rivals page is dense with tables, so a blunt rule floods false positives: start with the headline paragraphs and the README bottom line.
  • Score: 0.62 = 3 × 1 × 0.5 ÷ 3 × 1.25 × 1 (open rank 5 of 11); assumed: reach impact effort

P2: Measure the improvement loop against a plain 'what should we build next' prompt on two real repos

  • Labels: learning, CPO, stage:mvp
  • Where: docs/bench/ (a preregistration and a ledger) and the loop's reference.
  • Done when: Two real repos and at least 20 paired prompts judged by a rubric-free blind judge, the plain prompt as the control, spend capped and logged; the result and its interval are added to docs/perun-vs-rivals.md.
  • Verify: The preregistration commit precedes the run; the page states the interval and says tie when it contains 0.5.
  • Why: C-suite advice: the loop is the new capability, and a value claim needs a before-and-after measure, not an assertion. (vision: aligned)
  • Success metric: Win rate vs a plain prompt on a rubric-free judge: unmeasured now; 20 or more pairs; claim a win only if the lower bound is above 0.5.
  • Evidence: improvement-loop.md; perun-vs-rivals.md#h3-skill-routing-age…; bench
  • Risks: The judge is from the same vendor family and rubric bias is shown elsewhere on the rivals page: use a rubric-free judge and report that caveat.
  • Gate: owner: paid bench
  • Merged from: CPO lens: the visible-improvement bar; CMO lens: advice quality is the least measured of the five areas
  • Score: 0.60 = 2 × 2 × 0.5 ÷ 5 × 1.5 × 1 (open rank 6 of 11); assumed: reach impact effort

P3, no issue drafted until promoted: judge-independence-rerun, distribution-listing, real-host-smoke-tests, peer-survey-v2, product-skill-runnable-checks.

Generated by roadmap_rank.py from the input JSON: edit the input and re-run, never this text.


Agent suite roadmap (proposed / living)

Status: a design captured so nothing is lost — proposed, pending the owner decisions flagged in Decisions for the owner. Nothing here is built except items marked ✅. The deep-code-review skill name stays stable regardless of any repository/umbrella rename. Dated 2026-09-11.

BLUF

This repository is evolving from a portable code-review bar into a stage-aware agentic product-building suite: a "CEO" orchestrator plus opt-in specialist skills that support a solo or small builder across a product's whole life — code, infrastructure, docs, product/PMF, prioritization, discovery, growth, positioning, and business/ops — while preventing over-engineering and never fabricating. One epistemic spine holds every skill together. Navigation is by lenses over ~10 skills, not a sprawl of new ones. Build the spine and the lowest-risk skills first; add the advisory skills, on the tightest leash, last.

The epistemic spine (non-negotiable in every skill) — and the differentiator

  1. Frameworks + the user's own inputs + observable evidence. Apply named industry frameworks to the builder's actual context; never invent market/financial/competitor facts.
  2. Route the unknowable. When correctness depends on context the repo/agent does not hold (a market number, a legal/tax/securities question), hand it to real research or a licensed professional — never conclude on it.
  3. Stage-aware. Calibrate what to do and what not to do to lifecycle stage; the floor (security, secrets, auth, data-loss/backups) never relaxes.
  4. Surface uncertainty; skip rather than guess. Empty beats fabricated.

The market has converged on "AI cofounder / agent team" persona products; research found their limits language is near-empty. So the differentiator is not more personas — it is honesty, verification, and stage-awareness. That is the white space, and it is exactly this repo's existing thesis.

Architecture — lenses over ~10 skills (not 70)

A naive STAGE × AREA grid invites one skill per cell and detonates the repo's anti-duplication thesis. Instead, following the pattern deep-code-review already runs (fixed domains + an archetype → load map + a role overlay that "never adds or drops a domain"):

  • Keep ~10 skills; admit a new one only by the admission rule. A new skill is justified only when it cannot be a lens or reference over an existing skill, or when a safety boundary requires its own home (the reason the suite is 10, not 8 — business-ops keeps the call on your numbers separate from routing the binding part to a professional). Lenses and references are not skill rows; do not count them as such. STAGE and AREA are lenses that select and order existing content. The orchestrator routes (stage, area) → (skill, lens).
  • Navigation, two layers. Passive: each skill's frontmatter description is the routing key — disjoint "use when…", one home per skill, no duplication (Level-1 metadata is ~100 tokens/skill, so the orchestrator can hold the whole map and still route precisely). Active: the CEO reads (stage + need + artifact) and dispatches.
  • The suite map is a registry, not a bundle. It describes how to reach each skill; it never holds copies. One index serves agents and humans alike.
  • Topology by work type. Fan out (3–5 subagents, clear boundaries) only for read-mostly, decomposable work (research, audit, review); keep writing single-writer. Size effort to the project — a small/early project is one agent wearing several skill-hats, not a swarm burning tokens.

Stage model (shared) — ✅ shipped

prototype → mvp → growth → mature, calibrating demands (urgency), never a defect's intrinsic severity, and never the floor. Lives in deep-code-review (the first-response STAGE field + the Project stage section + the going-forward roadmap). Every advisory skill keys to it.

The suite (skills and lenses)

Skill Status Role
deep-code-review ✅ (stage-aware) The review bar; archetype/role/stage lenses; parallel-audit fan-out
agentic-delivery ✅ Gated G0–G10 build; roles as hats; one writer per worktree
idea-critic ✅ Pre-owner adversarial attack (the critic half of "critic + verifier")
communication-structure ✅ BLUF, one-ask human-facing messages
contribution ✅ (hardened — PR #26) Human-gated, privacy-safe upstream self-improvement
agent-readiness lens ✅ Is the repo agent-ready? (a role-coverage.md lens)
agentic-ceo (orchestrator) ✅ vendored (1.37.0) Suite registry + (stage, area) → (skill, lens) routing + effort-sizing + chaos playbook; defers general delivery/critique to agentic-delivery / idea-critic
infra-evolution-by-stage (ref) ✅ shipped 1.35.0 (#28→#51) When to add CI/staging/observability/k8s — extends the going-forward roadmap
docs-evolution-by-stage (ref) ✅ shipped 1.35.0 (#29→#51) Which docs at which stage; docs acquire normative force as the product matures
discovery / validation ✅ shipped in product-discovery (1.36.0) Design + interpret real user research; never fabricate findings
prioritization / focus ✅ shipped in product-discovery (1.36.0) Riskiest-assumption gate; what to build now vs. defer
product-strategy / PMF ✅ folded into product-discovery (1.39.0) — PMF read + Non-goals lens; standalone skill dropped per carve JTBD, PMF signals, what not to build
growth / analytics ✅ shipped in growth-analytics (1.38.0) Retention-first; North Star; what to instrument per stage
positioning / branding ✅ shipped in positioning (1.41.0) VPC → positioning → message house, as a hypothesis; thin guide; never fabricates TAM/competitor/quotes/clearance
business / ops ✅ shipped in business-ops (1.42.0) leads with the call on pricing/unit-economics numbers; routes the binding legal/tax/securities part to a professional; thin guide
CEO-under-pressure Researched → agentic-ceo (#36) Capture losslessly → reflect back → triage → one next action → tracked backlog → support by action (never "calm down")
responsible-AI / output safety Researched → build first among gaps (NEW skill) Govern the harm the product's own outputs do to end-users (NIST AI RMF lens)
further areas (gap-analysis) Researched → Further areas + issues regulated-triage, support-ops, op-readiness, privacy-by-design, decision-hygiene, billing — mostly lenses

Build order (research-backed)

  1. Critic + verifier — ✅ have (idea-critic, deep-code-review); assemble, don't rebuild.
  2. Discovery + PMF-gate — the top risk-reducer: a coding agent now builds the wrong thing faster, so the seat it does not fill (is this worth building?) is the highest leverage. New.
  3. Product-strategy / "what NOT to build" — governs the coding agent's backlog; prevents gold-plating.
  4. Analytics scoreboard + one primary metric — tells the builder whether the fast-built thing works; anchors every skill to one axis.
  5. Stage-aware CEO spine — extend agentic-ceo with stage/size-aware routing + the suite registry.
  • Shipped (1.35.0): the infra-evolution-by-stage and docs-evolution-by-stage references (extend the going-forward roadmap).
  • Defer / lightest touch: engineering advisor (already covered by the coding agent), design/UX (partly covered + product-ux-quality), legal/ops (one-time checklists + a tripwire), fundraising/hiring (stage-gated), positioning/branding + business/ops (highest fabrication-risk → thin "framework + the questions you must answer" guides, built last).

Sequencing rationale (and the pushback that produced it): do not big-bang a dozen skills — that is the exact over-engineering the suite exists to prevent, and it front-loads the riskiest (business/branding) skills before the pattern is proven. Build the spine + the lowest-fabrication-risk skills first, prove navigation end-to-end, then add the advisory skills with the tightest leash.

Per-area boundaries (frameworks by name; the hard line for each)

  • Discovery/validation: Mom Test, JTBD switch interviews, Teresa Torres continuous discovery, assumption/riskiest-assumption mapping, fake-door/concierge tests. Boundary: design and interpret research the user runs; never fabricate quotes, personas, market size, or a "validated" verdict; refuse to emit an "example finding" (it becomes fake evidence).
  • Product/PMF/prioritization: riskiest-assumption gate (name it, run the cheapest test before building — it produces a refusal); PMF read via the "very-disappointed" survey + a flattening retention cohort; ICE over RICE for a solo builder. Boundary: the skill owns the question, structure, and arithmetic; it routes for every input and never emits a scoring table with invented numbers.
  • Growth/analytics: retention-first (AARRR read bottom-up), a customer-value North Star (not vanity metrics), instrument only what answers a named question. Boundary: no fabricated benchmarks; route real figures to the user's analytics. Note: the product-analytics identifier rule inverts the ops-observability one (a stable per-user id is required for cohorts) — a seam to state, not restate.
  • Positioning/branding: Value Proposition Canvas → positioning → a message house → validation with real buyers; a "minimum viable brand" pre-PMF. Boundary: frameworks yield a hypothesis; the evidence comes from outside the model. Never fabricate TAM, competitor claims, quotes, outcomes, or trademark/domain clearance. Premature branding/scaling is a documented failure mode.
  • Infra-evolution-by-stage: infrastructure is earned, not provisioned — start with one boring, well-modularized deployable; add each layer (CI, staging, IaC, observability, containers, orchestration, service extraction) only on a concrete observable trigger. Boundary: the floor never relaxes; route the six business/ops unknowables (real SLA, regulatory class, revenue exposure, funding runway, team trajectory, customer concentration) to the owner. Complements infra-iac-containers.md (how-to-secure) with a when-to-add lens.
  • Docs-evolution-by-stage: docs acquire normative force as the product matures — one-pager (persuade) → design-doc/RFC (decide) → spec (contract, CI-enforced); trigger, not calendar. Agent-legibility = a small always-loaded router + depth routed with "read this when…". Complements docs-and-dx.md.
  • Business/ops (asymmetric boundary): The call on your numbers (pricing, unit economics): apply to the builder's own numbers with the formula and every assumption shown, flag thin data — but never turn arithmetic into a directive. The binding, regulation-dependent part (legal, tax, securities, employment classification, privacy compliance): route to a licensed professional, never conclude; fundraising (selling equity/SAFEs is a securities offering) is the boundary crossing. Standing disclaimer: educational information, not advice.

The product ↔ user ↔ agent loop ("growing together")

Two nested build-measure-learn loops. Loop 1 (the builder's product): the agent structures feedback and proposes experiments; the human decides. Loop 2 (the suite's own improvement): capture each interaction → evaluate with a split rubric (deterministic gates for the hard axes — boundary-compliance and non-fabrication, because a fluent wrong answer fools an LLM judge; LLM-judge, and a decorrelated one, for the soft axes only) → ratchet only through the human-gated, generalization-checked contribution flow. Quality only ratchets up: a regression gate stops backsliding, generalized eval cases raise the floor, and a human authorizes every landing. The fuel is verified eval cases and citations — not hoarded user data.

Operational runbook: How Perun improves itself makes this loop executable — the concrete 8 steps, the tool each one uses, and the built-vs-planned state of the split rubric.

Additive principles (generalized from field practice and research)

  • Registry, not a bundle — describe how to reach things; never duplicate them.
  • Store the query, not the answer; audit memory automatically — a stored status rots silently; store how to re-derive it, and flag memories that stopped being true.
  • Issue contract: Where / Done-when / Verify / Why — "Verify" is load-bearing; without it "done" is an opinion. Used for this repo's own backlog issues.
  • Unreadable must never look like empty (= the repo's UNVERIFIED); enforce rules in code, not prose (= contribution's kernel-paths.txt gate); deny by default. These validated recent work rather than changing it.

CEO under pressure (chaos playbook) — folds into agentic-ceo (#36)

When the owner floods the orchestrator with many rapid, conflicting asks under real pressure, the failure is to freelance (obey the last thing, drop the rest, act on colliding orders) or to patronize ("calm down"). Support is by action, not affect — six moves in order:

  1. Capture losslessly — every request/aside becomes a numbered logged item before any judgement; tag request vs. context; nothing filtered or merged.
  2. Reflect the full list back — the played-back numbered list is the acknowledgement; name any collision between items as its own finding.
  3. Triage to the vital few — a sequential first-hit-wins chain: blocks-others or irreversible → Now; failing on the live surface → Now; time-critical but reversible → Next; else → Held (visible, not dropped).
  4. One highest-leverage next action — the lead domino + a one-line why, offered as the existing A/B decision so the owner keeps final say.
  5. Hold the rest as a tracked backlog — each item has a visible state (in-flight / next / held / dropped-with-reason); WIP-limit to one primary action.
  6. Support by action — acknowledge the stakes as legitimate, show the list, name the move. Forbidden: "calm down"/"relax" (reactance + invalidation), toxic positivity, minimizing, narrating their feelings, "on it!" with nothing captured, silently absorbing conflicting orders.

Boundary: the agent is an organizer, not a therapist — the calm comes from the system being visibly under control. Capture-mode is not yes-mode: it still surfaces real disagreement (conflicts, quality-bar breaches, spread-thin mediocrity) in one line with the standard cited, then the owner decides. Frameworks (by name): GTD capture/clarify, incident command (single commander, unity-of-command, activity log), emergency-severity triage, WIP limits, "the ONE thing," psychological reactance, motivational-interviewing reflective listening.

Further areas (from gap-analysis)

The suite is strong on builder + product + code; the gap is the end-user runtime axis (what the product does to people once live) plus operational readiness for the bad day. Mostly lenses over existing skills, not new skills. Ranked:

  • G1 · Responsible-AI / product-output safety — NEW skill; govern harm from the product's own outputs (bias, hallucination-as-fact, missing AI-disclosure, deceptive patterns, unsafe automation). NIST AI RMF (GOVERN/MAP/MEASURE/MANAGE) lens; measure via the eval harness (domain J); human-in-the-loop on high-stakes outputs. The defining hole for an agentic-product suite — build first.
  • G2 · Regulated-domain obligation triage — small lens; a design-time decision tree (health? payments? minors? EU data? biometrics? consequential decisions?) that names the regime and routes before architecture hardens; re-runs on entering a new market. Cheapest build, earliest bite; unblocks G4/G5.
  • G3 · Support & feedback operations — lens; intake/triage/SLA + support→backlog.
  • G4 · Operational readiness — lens over agentic-delivery postmortem + M/Q: incident-response runbook, severity + comms + status page, break-glass access, credential/renewal inventory with dead-man reminders, solo-founder succession.
  • G5 · Privacy/compliance by design (product artifacts) — lens over Q: data inventory/ROPA, DPIA scaffold, consent-UX spec, subprocessor list; coupled to G2.
  • G6 · Decision hygiene for the builder's own calls — a routed reference extending idea-critic (one-/two-way doors, base rates, kill criteria set up front) — not a new skill.
  • G7 · Billing / monetization system correctness — a routed deep-code-review reference (metering, proration, dunning, tax, webhook idempotency, double-charge races) — mechanics of E/F/G/I applied to revenue.
  • Conditional: G8 trust & safety / abuse (multi-user products), G9 vendor business-risk, G10 accessibility-as-design (mostly redundant with domain P).
  • Do NOT build (redundant): cost / i18n / a11y-code-gate / supply-chain / data-integrity — already covered. Owned by planned skills (name in their charters): scope-defense → prioritization; data-as-moat → product-strategy; event-taxonomy → growth; pricing-strategy → business-ops; cloud right-sizing → infra-by-stage.
  • Boundary (all gaps): inventory / scaffold / checklist / red-team / cite-only- what-was-fetched; never fabricate a statute, deadline, CWE, or metric; route every legal/regime determination to a licensed professional. Legal specifics stay by-name / UNVERIFIED until fetched (as privacy-compliance.md already does). The NIST AI RMF core function names (Govern/Map/Measure/Manage) are confirmed (docs/standards-index.md); its AI RMF 1.0 / Generative AI Profile control specifics and GDPR/DPIA/COPPA/PCI/breach-clock specifics were not — fetch before citing.
  • Build first among gaps: G1, then G2 (unblocks G4/G5), then G3.

Decisions for the owner (open)

  1. Rename / umbrella. Defensible now that scope has widened (candidate: a Slavic-mythology theme, e.g. a "Perun" orchestrator, as a thin identity layer over functional skill names). Keep the deep-code-review skill name. Execute only after the spine is proven; owner picks the name; a migration checklist ships with it. GitHub auto-redirects the old repo URL, so the blast is mostly README / marketplace slug / plugin.json name.
  2. agentic-ceo: ✅ resolved — vendored into this repo (1.37.0, the suite conductor). Chosen over reference-global so the suite stays portable; coexists with any global skill of that name via project-scoping.
  3. Business-advisory skills: full "advisor" skills, or thin "framework + the questions you must answer yourself" guides? (Recommendation: thin guides — lowest fabrication surface.)
  4. Suite scale (#66): ✅ resolved — keep ~10 under a written admission rule (Architecture §). A new skill must fail the "can this be a lens/reference over an existing skill?" test or be a safety boundary; otherwise it lands as a lens/ref, and lenses/refs are not counted as skill rows. Applying it to the gap backlog: #43–#49 are lenses/references over existing skills; only #42 (product-output-safety / responsible-AI) is admitted as standalone skill #11 — output-harm governance is a distinct method from code-security and is a safety boundary. Residual risk: if #42 drafts thin, fold it as a deep-code-review lens rather than ship a standalone skill.

Status ledger (2026-09-11)

  • Suite complete + hardening — main @ 1.48.0 as of this snapshot (current release: see .claude/skills/deep-code-review/VERSION / CHANGELOG.md), 11 skills. Shipped: contribution overlay (1.31.0) · agent-readiness lens (1.32.0) · stage-aware review + going-forward roadmap (1.33.0) · contribution hardening (1.34.0; contribution 1.1.0) · infra/docs stage-evolution refs (1.35.0) · product-discovery (1.36.0, +Non-goals 1.39.0 → 1.1.0) · agentic-ceo conductor (1.37.0, +full registry 1.43.0 → 1.1.0) · growth-analytics (1.38.0) · positioning (1.41.0) · business-ops (1.42.0) · no-slop output contract (communication-structure → 1.2.0, at 1.40.0). Each new skill passed an independent reviewer pass before merge.
  • Post-build hardening (1.44–1.47): suite-enumeration gate (1.44.0, #62) · DCR fabrication-refusal eval (1.45.0, #65) · --recommend surfaces advisory overlays (1.46.0, #68) · README "which skill when" quickstart (#67) · offline eval-predicate discrimination gate (1.47.0, #61 slice 1) · growth-analytics benchmark binding (offline net) · product-output-safety = skill #11 (1.48.0, #42 — the one gap admitted standalone under the #66 rule).
  • Owner decisions: #39 (agentic-ceo) resolved — vendored; #40 (business-advisory shape) resolved — thin guides; #38 (rename) deferred until the suite is proven.
  • Post-build backlog (filed): #61 live eval harness · #62 suite-enumeration gate · #63 routing eval coverage · #64 eval-id anti-drift gate · #65 DCR fabrication eval · #66 [decision] suite scale — ✅ resolved (~10 + admission rule) · #67 README quickstart · #68 --recommend surfaces advisory skills · #69 dogfood the contribution loop. Gap-analysis skills: #42–#49.
  • Research complete: all 12 areas + CEO-under-pressure + gap-analysis. CEO-under- pressure shipped in agentic-ceo's chaos playbook.

This document is provisional and updated as skills ship. Decisions #39 (vendor), #40 (thin guides), and #66 (suite scale) are resolved and executed; #38 (rename) remains deferred. The suite is 10 skills under a written admission rule (Architecture §); #42 (product-output-safety) is the one admitted standalone candidate — skill #11.