frontier-scan

Skill: frontier-scan

The INSTANT-call version of the runbook in fdk/wiki/concepts/frontier-gap-scan.md. Goal: overstack is a COMPREHENSIVE frontier, not resting on wins. In place of the cloud routine (blocked because it needs a GitHub connection) — run manually at any time, human-in-the-loop.

WHAT

Purpose and context

Purpose: scout the agent-framework frontier over the last 30 days, score overstack across 8 axes (Ngang/Chớm/Thua), diff against the previous period, raise/update issues for gaps and update the report.
Trigger (when to use):
- "frontier scan", "scan competitors", "run scout", "weekly scout", "what are we losing to the world", "/frontier-scan".
- Periodically (intended weekly) or when news of a new framework/technique worth comparing comes up.
Non-goals: researching a general topic (that is /last30days); raising a single already-known issue (that is /raise-issue); coding a fix for a gap (the job of the receiving session, opened with /fdk).

Mental model

runbook + baseline kỳ trước → 5 WebSearch (4 cố định + 1 tự do) → verdict 8 trục so với CAPABILITIES → diff kỳ trước (tụt hạng / trục mới) → issue (raise mới hoặc comment cập nhật) → report + log.

Input and output contract

	Field	Required?	Meaning
In	runbook `fdk/wiki/concepts/frontier-gap-scan.md`	yes	8 axes + how to score
In	previous period's report `llmwiki/html/overstack-vs-world-30d.html`	yes	baseline for diff
In	`llmwiki/wiki/sources/ISSUES.md` + `fdk/CAPABILITIES.md`	yes	open issues + existing capabilities
Out	8-axis verdict + diff against previous period	yes	clearly state axes that dropped / new axes
Out	new issue (ledger + GH) or `gh issue comment`	per gap	not a duplicate of an existing issue
Out	updated report + 1 line `llmwiki/wiki/log.md`	yes	keep the old baseline, add a diff section for the new period

Rules and capabilities

RULE-01 (MUST): ONLY scan + raise/update + report. No coding, no self-merge, no raising an issue's status beyond open.
RULE-02 (MUST): Keep the 4 fixed scan queries across periods so the diff is meaningful; the fifth query is free-form.
RULE-03 (MUST): Query the wiki + ISSUES.md before raising (R7-f) — do not raise a duplicate of a gap that already has an issue.
RULE-04 (MUST): Definition of "win": no axis sits in Thua for more than 2 consecutive periods without an issue in flight.
RULE-05 (MUST): The repo forbids AI attribution in commits (R15) — if there is a commit, do not add Co-Authored-By.
RULE-06 (MUST): Touch only what the task requires — no opportunistic changes.
Capabilities: web search; read previous period's report + ledger; write report + log + ledger; write issues to the remote tracker.

Failure boundaries

Missing previous period report → cannot diff: clearly state this is the first baseline (partial), do not fabricate a previous-period verdict.
Remote tracker unreachable (not yet authenticated gh) → only write ledger + ISSUES.md (partial), notify user.
Gap uncertain whether an issue already exists → clarify/query first, do not raise a duplicate (RULE-03).

HOW

Main workflow

Step	Type	Inputs	Action	Outputs/exit	Failure/next
W01	deterministic	runbook, report, ISSUES.md	Read runbook + baseline	8 axes + previous verdict + open issues	missing report → partial (first baseline)
W02	effect	5 queries	Run 5 WebSearch (4 fixed + 1 free)	search results	—
W03	judgment	results + `fdk/CAPABILITIES.md`	Score 8 axes Level/Approaching/Losing	verdict	—
W04	judgment	new + old verdict	Diff vs previous period: rank drop / new axis	list of changes	—
W05	effect	gap	Raise (B01) or comment update (B02)	issue	duplicate → B02
W06	effect	verdict + diff	Update report + log	report + log	—
W07	deterministic	—	Stop, do not code fix	report	—

Step-by-step details (source of truth for W01–W07):

Read runbook + baseline: fdk/wiki/concepts/frontier-gap-scan.md and the most recent period report llmwiki/html/overstack-vs-world-30d.html. Load the list of 8 axes + previous verdict + open issues (llmwiki/wiki/sources/ISSUES.md, GH#9-13 + Ralph #15).
Scan (5 WebSearch): (a) Claude Code / agent framework updates this month; (b) AI agent memory + context engineering; (c) self-improving agents + harness eval; (d) agent skills marketplace + supply-chain security; (e) 1 query on a self-chosen hot topic. Keep the first 4 queries fixed to allow period-over-period comparison.
Compare: score 8 axes (Harness, Orchestration, Memory, Self-evolving skills, Knowledge/context, Eval/observability, Skill security, Scale) against fdk/CAPABILITIES.md → verdict Level/Approaching/Losing.
Diff vs previous period: which axis dropped in rank? which NEW axis appeared? This is the most important signal — state it clearly.
Raise / update (R7-f query first to avoid duplicates):
- NEW gap or WORSENED gap with no issue yet → /raise-issue (ledger draft + ISSUES.md line + gh issue create --assignee Rheinmir, ledger backlink body, add "Reference repo/paper" section).
- Gap already has an issue → ONLY gh issue comment update evidence, do NOT raise a duplicate.
Report + log: update llmwiki/html/overstack-vs-world-30d.html (keep the old baseline, add the new-cycle diff section) + append 1 line of llmwiki/wiki/log.md.
Stop, do NOT code fix: this skill only scouts + raises. Executing the gap is the job of the receiving session (opened with /fdk).

Branches

ID	Kind	Guard	Behavior	Skip / failure	Rejoin
B01	conditional_required	NEW or WORSENED gap, no issue yet	`/raise-issue` (ledger + ISSUES.md + `gh issue create`)	—	W06
B02	conditional_required	gap already has an issue	only `gh issue comment` update evidence	—	W06

Validation and stopping

Every new issue must pass query R7-f (no duplicates); the report keeps the old baseline. The 8-axis verdict is a judgment — it needs a human reader's approval. Stop after W07, do not move on to code.

Examples

Positive: "/frontier-scan" this week → the Memory axis drops from Ngang → Chớm, no issue yet → /raise-issue + ISSUES.md line; the report adds the new-cycle diff section + 1 log line.
Boundary/failure: the Skill security axis is still Thua but already has an open issue in ISSUES.md → only gh issue comment new evidence, do NOT raise a second issue.