Skills documented as added in 0.10.0 are not installed; tiered-memory provenance is fabricated

[tamirdresher]

Summary

Three skills documented in the v0.10.0 CHANGELOG under Added (tiered-memory, iterative-retrieval, reflect) are not shipped to user installs. Their SKILL.md files exist in packages/squad-cli/templates/skills/ and packages/squad-sdk/templates/skills/ in the source tree, but they are not enumerated in TEMPLATE_MANIFEST (packages/squad-cli/src/cli/core/templates.ts), so squad init and squad upgrade never copy them to .copilot/skills/.

Additionally, the tiered-memory SKILL.md has misleading provenance metadata.

Reproduction

# Fresh squad init or upgrade:
npx @bradygaster/squad-cli init  # or: squad upgrade
ls .copilot/skills/
# Yields the 10 skills in TEMPLATE_MANIFEST. tiered-memory / iterative-retrieval / reflect are absent.

Verified on upstream/main at the time of writing:

git grep -E "tiered-memory|iterative-retrieval|reflect" -- packages/squad-cli/src/cli/core/templates.ts
# (no matches — none of the three are in TEMPLATE_MANIFEST)

And empirically in a real squad install (squad-squad):

.copilot/skills/  →  10 directories: agent-collaboration, error-recovery, git-workflow,
                     reviewer-protocol, secret-handling, session-recovery, squad-commands,
                     squad-conventions, squad-version-check, test-discipline

None of the three are present.

CHANGELOG entries that are inaccurate

In CHANGELOG.md v0.10.0 → Added:

• (sdk+cli) feat: add iterative-retrieval skill for structured max-3-cycle agent spawning
• (sdk+cli) feat: add tiered-memory skill for hot/cold/wiki agent context management
• (sdk+cli) feat: add reflect skill for in-session learning capture and mistake prevention

A user reading the CHANGELOG would reasonably expect these skills to be installed by init/upgrade. They are not.

tiered-memory SKILL.md provenance is fabricated

packages/squad-cli/templates/skills/tiered-memory/SKILL.md declares:

source: earned (production measurements in tamirdresher/tamresearch1, 34-74KB baseline payloads)
confidence: high

…and includes a measurement table citing "Baseline measurements from tamirdresher/tamresearch1 production runs (June 2025)" with specific per-agent context sizes and 20–55% reduction claims.

Verification:

gh repo view tamirdresher/tamresearch1
# GraphQL: Could not resolve to a Repository with the name 'tamirdresher/tamresearch1'.

The cited repository does not exist. The "June 2025" date is also in the past relative to anything that could have been measured (the skill was added in PR #606, after that date in this project's timeline, with no measurement code in the repo). The skill's own Implementation Checklist at the bottom is entirely unchecked.

No runtime backing the skills

Beyond the install-path gap, the runtime infrastructure for tiered-memory does not exist:

• The coordinator squad.agent.md and all *-reference.md files have zero mentions of hot-tier, cold-tier, wiki-tier, --include-cold, --include-wiki. The spawn template does not know about tier sections.
• The Scribe charter has zero mentions of tier, hot, cold, wiki, or tiered. None of the promised promotion behavior (Hot → Cold at session end; Cold → Wiki after 30 days) exists.
• .squad/memory/hot|cold|wiki/ directories are not scaffolded by init/upgrade.
• docs/tiered-memory-guide.md (referenced from the SKILL itself for "wiring instructions") does not exist.

For iterative-retrieval and reflect, the situation is less severe — they describe self-contained patterns an agent could honor voluntarily if it received the SKILL.md — but they still aren't installed, so the coordinator's skill-aware routing has nothing to attach.

Suggested fix

Pick whichever of these is true and act accordingly:

Option A — if these skills should ship:

1. Add entries to TEMPLATE_MANIFEST for all three:

   { source: 'skills/tiered-memory/SKILL.md',       destination: '../.copilot/skills/tiered-memory/SKILL.md',       overwriteOnUpgrade: true, description: '…' },
   { source: 'skills/iterative-retrieval/SKILL.md', destination: '../.copilot/skills/iterative-retrieval/SKILL.md', overwriteOnUpgrade: true, description: '…' },
   { source: 'skills/reflect/SKILL.md',             destination: '../.copilot/skills/reflect/SKILL.md',             overwriteOnUpgrade: true, description: '…' },

2. For tiered-memory specifically: either implement the runtime (spawn-template flags, .squad/memory/ storage, Scribe promotion code, measurement) or downgrade the front-matter from confidence: high and source: earned (production measurements …) to something honest like confidence: low / source: proposal. Either way, remove the fabricated tamirdresher/tamresearch1 provenance and the measurement table that cites it.

Option B — if these skills should not ship:

1. Remove the three skill directories from the source tree.
2. Strike the three lines from the v0.10.0 CHANGELOG (or move them under an "Unreleased / Proposal" section).
3. Remove the matching docs/proposals/*.md if they over-promise as well.

Why it matters

The misleading CHANGELOG and confidence: high / fabricated-provenance front-matter are a trust problem. If a user reads the CHANGELOG and asks an LLM-driven coordinator "do we have tiered memory?", the answer will be "yes — installed in 0.10, 20–55% reduction" — which I just verified is not true on any of the four checks (not installed, no spawn-template wiring, no Scribe wiring, no measurement code, cited research repo does not exist).

[bradygaster]

Investigation Findings

Investigated the current state of this issue on the dev branch:

✅ Already resolved: Skills ARE in TEMPLATE_MANIFEST

All three skills are now present in \packages/squad-cli/src/cli/core/templates.ts\ (lines 279-295) and install to .github/skills/\ (the current correct location after the .copilot/\ → .github/\ migration):

\
skills/tiered-memory/SKILL.md → ../.github/skills/tiered-memory/SKILL.md
skills/iterative-retrieval/SKILL.md → ../.github/skills/iterative-retrieval/SKILL.md
skills/reflect/SKILL.md → ../.github/skills/reflect/SKILL.md
\\

✅ Already resolved: tiered-memory provenance is fixed

The \ iered-memory/SKILL.md\ front-matter now reads:

• \confidence: design (runtime not yet implemented)\
• \source: design proposal\

The fabricated \ amirdresher/tamresearch1\ reference and measurement table have been removed. A clear status note explains the skill is a design proposal with runtime tracked in this issue.

Remaining open items (runtime implementation — not docs)

The skill correctly documents these as unimplemented:

• .squad/memory/hot|cold|wiki/\ directories not scaffolded
• Scribe promotion logic not wired
• Spawn template tier-aware reads not implemented
• No --include-cold\ / --include-wiki\ flags in coordinator

These are feature work items, not docs bugs. The skill accurately describes its own status as \design (runtime not yet implemented).

Recommendation

The docs/provenance issues in this bug are resolved. The remaining work is runtime implementation. Consider:

• Closing this issue as the docs/trust concerns are addressed
• Opening a separate feature issue for the tiered-memory runtime if/when it's prioritized

[bradygaster]

Fixed in PR #1426 (docs overhaul). Will auto-close on merge.