[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7ENA8VTW7REFQ01MHSGGXZK":3,"skill-same:01M7ENA8VTW7REFQ01MHSGGXZK":46},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M7ENA8VTW7REFQ01MHSGGXZK","learn",null,"Post-engagement knowledge harvest AND harness retrospective - after a box\u002Fbugbounty\u002Fpentest\u002FCTF is completed, first diff how the engagement was EXECUTED against the skills\u002Fhooks that governed it (what discipline was skipped) and improve the harness, then sweep the whole engagement for GENERIC reusable knowledge NOT already in wiki\u002F and land it via the leak-gated stage->promote pipeline. Use at close-out or when asked to \"extract learnings\", \"what did we learn\", \"harvest lessons into wiki\", \"distill this engagement\", \"post-mortem into the wiki\", \"what did we do wrong\", \"improve the harness from this box\".","lza6\u002FArmor-piercing-system-prompts","lza6","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F128880206?v=4","01M7EB3K7WPX5T28EJ2NVHE6EG","Chinese-tingfeng","01M7EB3K97E1KQ16QWKJMFNGX8",11,"MIT","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F04-安全渗透\u002Flearn-TORCH\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: learn\ndescription: Post-engagement knowledge harvest AND harness retrospective - after a box\u002Fbugbounty\u002Fpentest\u002FCTF is completed, first diff how the engagement was EXECUTED against the skills\u002Fhooks that governed it (what discipline was skipped) and improve the harness, then sweep the whole engagement for GENERIC reusable knowledge NOT already in wiki\u002F and land it via the leak-gated stage->promote pipeline. Use at close-out or when asked to \"extract learnings\", \"what did we learn\", \"harvest lessons into wiki\", \"distill this engagement\", \"post-mortem into the wiki\", \"what did we do wrong\", \"improve the harness from this box\".\n---\n\n# Learn: harness retrospective + wiki knowledge harvest\n\nThe close-out retrospective, in two phases:\n- **Phase 0 (process):** diff how the engagement was EXECUTED against the discipline the\n  skills\u002Fhooks prescribe, find where the operator (you) or the harness drifted, and fix the\n  harness so the next box does not repeat it. This is the \"improve the harness on it\" half.\n- **Phases 1-7 (knowledge):** read the WHOLE finished engagement, diff it against the existing\n  wiki, and promote only the DELTA as durable, generic knowledge. The safety net that catches\n  everything not captured live during the box.\n\nThe client-data boundary is enforced by code, not by this prose: engagement-derived\ncontent reaches `wiki\u002F` ONLY through `wiki-stage.py` -> `wiki-promote.py`, which runs\n`check-leaks.sh` and fails closed. This skill READS the engagement and WRITES only\ngeneric knowledge; it never edits the engagement's own files.\n\n## When it fires\n- At close-out: run `Skill(learn)` once the engagement is marked\n  `## STATUS: SOLVED` (OWNED\u002FROOTED\u002FCOMPLETE) AND its `walkthrough.md` is assembled\n  (the CLAUDE.md execution-loop close-out step).\n  It self-clears the moment this skill writes `\u003Ceng>\u002F.learn-done`.\n- On demand: \"extract learnings\", \"what did we learn\", \"harvest lessons\", \"distill\",\n  \"post-mortem into the wiki\" - the same steps apply.\n\n## Phase 0: Process retrospective (harness self-improvement) - DO FIRST\n\nBefore harvesting knowledge, ask: **did we execute the way the skills\u002Fhooks told us to, and\nwhere we did not, whose fault is it - mine, or the harness's for not catching me?** The point is\nnot self-flagellation; it is to turn each drift into a concrete harness change so the next\nengagement cannot repeat it.\n\n### 0a. Diff execution against discipline\nRe-read the skills that governed this engagement (`ctf-box`, the `hunt-*` used, the CLAUDE.md\nexecution loop) and check each mandated step against what actually happened on disk:\n\n| Discipline it prescribes | How to check it was (not) done | Common drift |\n|---|---|---|\n| Recon tooling complete (nmap AND ffuf AND nuclei-read for web) | `ls targets\u002F$ENG\u002Frecon\u002F*.png`, `.recon-tools` marker, tmux window names, `log.md` | ffuf never run; nuclei launched but output never read |\n| Screenshot EVERY finding as it lands | count deliberate `poc\u002F*.png` vs findings in `Killchain.md`; `.screenshot-nudged` | shots only at the very end; transient states lost |\n| Wiki-first before exploiting each fingerprinted service | wiki queries in transcript \u002F `log.md` | jumped to exploitation from memory |\n| A hook nudge fired -> was it acted on? | grep the transcript for a nudge (e.g. \"switch to ffuf\") whose action never followed | nudge ignored under momentum |\n| State-first \u002F capture-as-you-go | `state.md`\u002F`loot.md` updated mid-box vs all-at-end | prose-in-chat lost |\n\n**If the engagement ran under the campaign driver** (`bb\u002Fpt\u002Fctf-workflow`), pull its own numbers -\nthey say precisely where the driver fought the operator, no guessing:\n\n```\npython3 scripts\u002Fcampaign.py --eng \u003CENG> ledger --json\n```\n\n`drift` = network calls the driver never emitted (improvisation off the board); `dry_rounds` \u002F\n`reframe_lenses_used` = how hard the campaign had to reframe to find anything; `paused_hosts` = hosts\nthat banned us; `board.dead` vs `board.closed` = the exhaustion-to-finding ratio. A high drift or a\nlong lens list is the retro's starting point.\n\nFor each drift, name the **root cause** honestly: a *skill* that under-specified, a *hook* that\nshould have fired a reflex but did not (or fired and was ignorable), or a *me* failure the harness\nhad no mechanism to catch.\n\n### 0b. Turn each drift into a harness change\nRoute each finding to its fix TARGET and make the change (small+clear -> apply now with the\nnormal edit+test loop; larger -> record as a proposed item for operator approval):\n- **Reflex gap** (a capture\u002Fcoverage signal the harness could have caught) -> add\u002Fextend a hook\n  in `skills\u002Fhooks\u002F` (stay within \"reflexes = capture\u002Froute\", NOT methodology), with a test in\n  `tests\u002F`. Run `python3 -m pytest tests\u002F -q` green before moving on.\n- **Discipline under-specified** -> tighten the governing skill (`ctf-box`\u002F`hunt-*`): make the\n  step an explicit ordered checklist item, not buried prose. Skill bodies carry lesson PROSE +\n  `[[wikilink]]` ONLY - never inline runnable exploit\u002Frootkit\u002FPoC code (magic-signal dumps,\n  syscall PoCs, CVE exploits, reversing walks). That code is the KNOWLEDGE half: promote it to a\n  `wiki\u002F` page through the Phase 5-6 gate and point the skill at it with a one-line \"try X (see\n  [[page]])\". A skill body loads wholesale into model\u002FAPI context on EVERY invocation, so raw\n  offensive code there trips the Claude API safety classifier (real: ctf-box's LKM-rootkit +\n  CVE-PoC lessons got flagged on load, forcing a manual purge) and bloats context. Prose that\n  says WHAT to try in WHAT order belongs in the skill; the code that does it belongs in the wiki.\n- **Script\u002Fanalyzer gap** -> extend `scripts\u002F` (coverage, next_move, find-lint) to surface it.\n\n### 0c. Log the retrospective (generic, tracked)\nAppend a dated section to `docs\u002Fsuperpowers\u002Fharness-retro.md` (create if missing): the drifts\nfound, their root cause, and the harness change made or proposed. This doc is about the HARNESS,\nso it is generic and tracked - describe failures generically (\"on a web box ffuf was skipped\"),\nnever with client host\u002FIP\u002Fcred.\n\n**Naming the engagement in this tracked file:** use its dir name ONLY when that name is a neutral\ncodename (`\u003Cplatform>_\u003Cboxname>`). Many engagement dirs are named after the target itself (a host, a\nproduct, a client short-name) - writing one of those into `harness-retro.md` publishes a client\nmarker, and `check-leaks.sh` derives its marker list from `targets\u002F` dir names, so it WILL fail the\ngate. When the dir name is target-derived, describe the engagement generically instead (\"a Symfony\nAPI target\", \"a single-host web bug-bounty campaign\") and name nothing. Run `bash\nscripts\u002Fcheck-leaks.sh` after writing the retro; a FAIL here almost always means the dir name leaked.\n\n### 0d. Fill the engagement's agent eval (per-engagement, gitignored)\n`targets\u002F$ENG\u002Feval.md` has two halves - auto (real data) and judgement (you). Scaffolded upfront\nfrom `_eval.md` for pentest\u002Fbugbounty; for ctf it self-creates the moment the next step\n(`eval_metrics.py --write`, already run once by close-out.py at SOLVED) writes to it, so it may not\nexist yet if you are running this by hand before close-out fired:\n\n**Auto first - REAL numbers, no estimation:**\n```bash\npython3 scripts\u002Feval_metrics.py $ENG --write\n```\nThis injects a `## Metrics (auto)` block from the hook telemetry (`.events.jsonl`) + the session\ntranscript recorded in `.metrics.json`: exact skill \u002F hook \u002F tool call counts, auto-detected drift\nsignals (every scope-guard block, state-discipline nudge, leak-warn), the start->finish time delta +\nidle-filtered active time, and the box's token usage (windowed to its active period). NEVER hand-type\nthese - the transcript is ground truth (a real box's true output was ~12x an earlier hand-estimate).\n\n**Then the judgement half - only you can write it:** in the same file, fill the Drift moments (what\nthe auto drift-count can't say: WHY each drift happened + the one-line fix), What went right, and the\nScores. The auto block counts drift; you narrate its cause. Dead-ends come from `Deadends.md`.\nUnlike the retro (generic, tracked), eval.md is per-engagement and may name specifics (gitignored `targets\u002F`).\n\nOnly after Phase 0 is done, proceed to the knowledge harvest below.\n\n## Steps (Phases 1-7: wiki knowledge harvest)\n\n### 1. Resolve the engagement\n```bash\nENG=$(cat targets\u002Factive.md)\nTYPE=$(grep -m1 engagement_type targets\u002F$ENG\u002Fstate.md | cut -d: -f2 | tr -d ' ')\n```\nConfirm close-out (`## STATUS:` heading in `state.md`) or that the operator asked\nexplicitly. Do not harvest an in-flight engagement unless asked.\n\n### 2. Inventory what the engagement taught\nRead the full engagement and list every candidate GENERIC lesson:\n- `state.md` - tech\u002Fservice\u002Fversion that mattered and how it was handled.\n- `loot.md` - default or vendor-known creds (NOT client-set passwords).\n- `Killchain.md` + `walkthrough.md` - the chain that actually worked, with exact commands.\n- `Deadends.md` - what failed. Negative knowledge is reusable: a bypass that does NOT\n  work on tech X, a default cred changed in vendor version Y, a false-positive pattern.\n- `Vuln-index.md` \u002F `Vulns\u002F` - findings and their reusable exploitation technique.\n- `Approach.md`, `log.md` - anything else non-obvious that recurs.\n\nA candidate is worth harvesting only if it is REUSABLE on the next engagement. Skip\none-off client trivia.\n\n### 3. Generalize + strip client specifics\nRewrite each candidate to its generic form: product + technique\u002Fcred\u002Fendpoint +\nimpact. Drop every client host, IP, domain, and client-set credential value. Client\nspecifics stay under `targets\u002F\u003Ceng>\u002F`. (The leak gate will refuse anything that slips\nthrough, but strip up front - do not lean on the gate.)\n\n**Compact to the DELTA, not a re-explanation.** A staged candidate is the MINIMAL new reusable\nprimitive + its one gotcha (a technique delta is typically \u003C ~15 lines), NOT a re-teach of what the\ntarget page already covers. Before promote, re-read the target section (Phase 4) and cut every line\nalready there -- keep ONLY what the wiki did not have. A verbose dump that repeats known material is\na failed harvest even if it promotes.\n\n### 4. Dedup against the wiki (the skip rule)\nFor each generic candidate, search the wiki first:\n```\nmcp__wiki-search__qmd_query   # semantic: concept \u002F technique \u002F intent\nmcp__wiki-search__qmd_search  # keyword: exact tool name, CVE id, payload string\n```\nFind the home page (one class = one page). Read only its frontmatter and the relevant\nsection. If the technique\u002Fpayload\u002Fcred is ALREADY covered there, SKIP it - wiki has it.\nKeep only the delta. This step is the whole point: \"extract stuff we haven't had yet\".\n\n### 5. Route each survivor through stage\nNever hand-edit `wiki\u002F` with engagement-derived content. Stage it:\n```bash\n# default \u002F vendor-known credential -> the cred cheatsheet\npython3 scripts\u002Fwiki-stage.py --kind default-cred --slug \u003Cproduct>-default \\\n  --body '| \u003Cproduct> | \u003Cversion> | \u003Cuser> | \u003Cpass> | observed | \u003Cgeneric note> |'\n\n# reusable request \u002F payload pattern -> the api-request cheatsheet\npython3 scripts\u002Fwiki-stage.py --kind api-pattern --slug \u003Cproduct>-\u003Cendpoint> \\\n  --body '| \u003Cproduct> | \u003Cendpoint> | \u003Cmethod> | \u003Crequest\u002Fpayload> | \u003Cauth> | \u003Creveals> |'\n\n# technique \u002F bypass \u002F tool-gotcha that ENRICHES an existing page\npython3 scripts\u002Fwiki-stage.py --kind technique --slug \u003Cslug> \\\n  --target-page techniques\u002F\u003Carea>\u002F\u003Cpage>.md\n# then edit targets\u002F$ENG\u002Fwiki-candidates\u002F\u003Cslug>.md to APPEND the generic '## Heading' body\n# (append below the staged frontmatter -- do NOT overwrite the file, or you drop `status: pending`\n#  and `wiki-promote --list` will show nothing to promote)\n```\n\n**Genuinely new class with no home page:** `wiki-promote` merges into an EXISTING\npage and skips a missing target. So first create a CONTENT-FREE generic scaffold\n(frontmatter per `docs\u002Fpage-types.md` + the section headings only, zero engagement\ndata), then stage the body against it so the substance still arrives through the gate:\n```bash\n# 1. scaffold wiki\u002Ftechniques\u002F\u003Carea>\u002F\u003Cslug>.md: frontmatter + empty section headings\n# 2. stage the generic body:\npython3 scripts\u002Fwiki-stage.py --kind technique --slug \u003Cslug> \\\n  --target-page techniques\u002F\u003Carea>\u002F\u003Cslug>.md\n```\nFor an external source that also informs the lesson (a CVE writeup, an advisory), hand\nthat part to `Skill(research-ingest)` rather than duplicating it here.\n\n### 6. Promote through the leak gate\n```bash\npython3 scripts\u002Fwiki-promote.py --list            # review pending candidates\npython3 scripts\u002Fwiki-promote.py --review \u003Cslug>    # read one in full\npython3 scripts\u002Fwiki-promote.py --promote all      # leak-checked merge + re-index\n```\nPromote runs `check-leaks.sh` on each body and refuses (writes nothing) on a client\nmarker. Report what promoted, what was refused, and why. `--promote` also appends a\ngeneric delta-yield tally (date, engagement_type, count) to `docs\u002Fwiki-delta-log.md`;\na zero-yield harvest is worth a sanity check (was the sweep actually run, or is the wiki\nalready saturated for this box).\n\n### 7. Re-index, lint, self-clear\n```bash\npython3 scripts\u002Flint-wiki.py -q        # must be clean (broken links, stale index)\n# only if a NEW page was created:\npython3 scripts\u002Fgen_index.py && python3 scripts\u002Fbuild_moc.py && qmd update\n# GATE: do NOT self-clear with eval.md's judgement half blank (Phase 0d). This exits 1 and\n# refuses the marker until Drift moments \u002F What went right \u002F Score are filled -- learn self-cleared\n# with the human half empty on a real box, so this is enforced, not advisory.\npython3 scripts\u002Feval_metrics.py $ENG --check-judgement || { echo \"fill eval.md judgement half first\"; exit 1; }\ntouch targets\u002F$ENG\u002F.learn-done         # marks this engagement's learn pass done\n```\nLog one GENERIC line to `session\u002Flog.md` (e.g. \"learn: promoted 3 -> jwt-attacks,\ndefault-credentials\") and an audit line to `targets\u002F$ENG\u002Flog.md`. Never put a client\nhost\u002FIP\u002Fdomain in either.\n\n## Guardrails\n- **Generic only, gate-enforced.** All engagement-derived writes go stage -> promote;\n  the leak check is the code boundary. A new page is a content-free scaffold; its\n  substance still comes through the gate.\n- **Dedup first.** Enrich the existing page; do not create a second page for a class\n  wiki already covers. `qmd` before staging, every time.\n- **No fabrication.** Harvest only what the engagement files actually record. If a\n  lesson is not backed by the state\u002Floot\u002FKillchain\u002Fwalkthrough\u002FDeadends record, drop it.\n- **Read-only on the engagement.** Phase 0 may edit HARNESS files (`skills\u002F`, `skills\u002Fhooks\u002F`,\n  `scripts\u002F`, `tests\u002F`, `docs\u002Fsuperpowers\u002Fharness-retro.md`) and Phases 1-7 write to `wiki\u002F`\n  (via the gate) and the `.learn-done` marker - but NEVER the engagement's own findings\u002Fnarrative.\n- **No offensive code in a skill body.** A tightened skill (`ctf-box`\u002F`hunt-*`) gets lesson prose +\n  `[[wikilinks]]` only. Runnable exploit\u002Frootkit\u002FPoC code lives in `wiki\u002F` (through the Phase 5-6\n  gate) and is referenced by link, never inlined - the skill loads into API context wholesale and\n  the classifier flags raw offensive code (it has). Reversing walks, syscall PoCs, magic-signal\n  dumps, CVE exploits = wiki page + a one-line skill pointer.\n- **Phase 0 changes ship green.** Any hook\u002Fscript edit lands with a test and a passing\n  `python3 -m pytest tests\u002F -q`; do not leave the harness broken by an improvement.\n- **Retro log is generic + tracked.** `harness-retro.md` describes harness failures generically\n  (no client host\u002FIP\u002Fcred); the leak boundary still applies.\n\nReport: **Phase 0** - drifts found, root cause each, harness changes made vs proposed, retro-log\npath, test status. **Phases 1-7** - candidates found, skipped-as-already-in-wiki, staged,\npromoted, refused, pages touched (and any new page created), lint status.\n","# Learn: harness retrospective + wiki knowledge harvest\n\nThe close-out retrospective, in two phases:\n- **Phase 0 (process):** diff how the engagement was EXECUTED against the discipline the\n  skills\u002Fhooks prescribe, find where the operator (you) or the harness drifted, and fix the\n  harness so the next box does not repeat it. This is the \"improve the harness on it\" half.\n- **Phases 1-7 (knowledge):** read the WHOLE finished engagement, diff it against the existing\n  wiki, and promote only the DELTA as durable, generic knowledge. The safety net that catches\n  everything not captured live during the box.\n\nThe client-data boundary is enforced by code, not by this prose: engagement-derived\ncontent reaches `wiki\u002F` ONLY through `wiki-stage.py` -> `wiki-promote.py`, which runs\n`check-leaks.sh` and fails closed. This skill READS the engagement and WRITES only\ngeneric knowledge; it never edits the engagement's own files.\n\n## When it fires\n- At close-out: run `Skill(learn)` once the engagement is marked\n  `## STATUS: SOLVED` (OWNED\u002FROOTED\u002FCOMPLETE) AND its `walkthrough.md` is assembled\n  (the CLAUDE.md execution-loop close-out step).\n  It self-clears the moment this skill writes `\u003Ceng>\u002F.learn-done`.\n- On demand: \"extract learnings\", \"what did we learn\", \"harvest lessons\", \"distill\",\n  \"post-mortem into the wiki\" - the same steps apply.\n\n## Phase 0: Process retrospective (harness self-improvement) - DO FIRST\n\nBefore harvesting knowledge, ask: **did we execute the way the skills\u002Fhooks told us to, and\nwhere we did not, whose fault is it - mine, or the harness's for not catching me?** The point is\nnot self-flagellation; it is to turn each drift into a concrete harness change so the next\nengagement cannot repeat it.\n\n### 0a. Diff execution against discipline\nRe-read the skills that governed this engagement (`ctf-box`, the `hunt-*` used, the CLAUDE.md\nexecution loop) and check each mandated step against what actually happened on disk:\n\n| Discipline it prescribes | How to check it was (not) done | Common drift |\n|---|---|---|\n| Recon tooling complete (nmap AND ffuf AND nuclei-read for web) | `ls targets\u002F$ENG\u002Frecon\u002F*.png`, `.recon-tools` marker, tmux window names, `log.md` | ffuf never run; nuclei launched but output never read |\n| Screenshot EVERY finding as it lands | count deliberate `poc\u002F*.png` vs findings in `Killchain.md`; `.screenshot-nudged` | shots only at the very end; transient states lost |\n| Wiki-first before exploiting each fingerprinted service | wiki queries in transcript \u002F `log.md` | jumped to exploitation from memory |\n| A hook nudge fired -> was it acted on? | grep the transcript for a nudge (e.g. \"switch to ffuf\") whose action never followed | nudge ignored under momentum |\n| State-first \u002F capture-as-you-go | `state.md`\u002F`loot.md` updated mid-box vs all-at-end | prose-in-chat lost |\n\n**If the engagement ran under the campaign driver** (`bb\u002Fpt\u002Fctf-workflow`), pull its own numbers -\nthey say precisely where the driver fought the operator, no guessing:\n\n```\npython3 scripts\u002Fcampaign.py --eng \u003CENG> ledger --json\n```\n\n`drift` = network calls the driver never emitted (improvisation off the board); `dry_rounds` \u002F\n`reframe_lenses_used` = how hard the campaign had to reframe to find anything; `paused_hosts` = hosts\nthat banned us; `board.dead` vs `board.closed` = the exhaustion-to-finding ratio. A high drift or a\nlong lens list is the retro's starting point.\n\nFor each drift, name the **root cause** honestly: a *skill* that under-specified, a *hook* that\nshould have fired a reflex but did not (or fired and was ignorable), or a *me* failure the harness\nhad no mechanism to catch.\n\n### 0b. Turn each drift into a harness change\nRoute each finding to its fix TARGET and make the change (small+clear -> apply now with the\nnormal edit+test loop; larger -> record as a proposed item for operator approval):\n- **Reflex gap** (a capture\u002Fcoverage signal the harness could have caught) -> add\u002Fextend a hook\n  in `skills\u002Fhooks\u002F` (stay within \"reflexes = capture\u002Froute\", NOT methodology), with a test in\n  `tests\u002F`. Run `python3 -m pytest tests\u002F -q` green before moving on.\n- **Discipline under-specified** -> tighten the governing skill (`ctf-box`\u002F`hunt-*`): make the\n  step an explicit ordered checklist item, not buried prose. Skill bodies carry lesson PROSE +\n  `[[wikilink]]` ONLY - never inline runnable exploit\u002Frootkit\u002FPoC code (magic-signal dumps,\n  syscall PoCs, CVE exploits, reversing walks). That code is the KNOWLEDGE half: promote it to a\n  `wiki\u002F` page through the Phase 5-6 gate and point the skill at it with a one-line \"try X (see\n  [[page]])\". A skill body loads wholesale into model\u002FAPI context on EVERY invocation, so raw\n  offensive code there trips the Claude API safety classifier (real: ctf-box's LKM-rootkit +\n  CVE-PoC lessons got flagged on load, forcing a manual purge) and bloats context. Prose that\n  says WHAT to try in WHAT order belongs in the skill; the code that does it belongs in the wiki.\n- **Script\u002Fanalyzer gap** -> extend `scripts\u002F` (coverage, next_move, find-lint) to surface it.\n\n### 0c. Log the retrospective (generic, tracked)\nAppend a dated section to `docs\u002Fsuperpowers\u002Fharness-retro.md` (create if missing): the drifts\nfound, their root cause, and the harness change made or proposed. This doc is about the HARNESS,\nso it is generic and tracked - describe failures generically (\"on a web box ffuf was skipped\"),\nnever with client host\u002FIP\u002Fcred.\n\n**Naming the engagement in this tracked file:** use its dir name ONLY when that name is a neutral\ncodename (`\u003Cplatform>_\u003Cboxname>`). Many engagement dirs are named after the target itself (a host, a\nproduct, a client short-name) - writing one of those into `harness-retro.md` publishes a client\nmarker, and `check-leaks.sh` derives its marker list from `targets\u002F` dir names, so it WILL fail the\ngate. When the dir name is target-derived, describe the engagement generically instead (\"a Symfony\nAPI target\", \"a single-host web bug-bounty campaign\") and name nothing. Run `bash\nscripts\u002Fcheck-leaks.sh` after writing the retro; a FAIL here almost always means the dir name leaked.\n\n### 0d. Fill the engagement's agent eval (per-engagement, gitignored)\n`targets\u002F$ENG\u002Feval.md` has two halves - auto (real data) and judgement (you). Scaffolded upfront\nfrom `_eval.md` for pentest\u002Fbugbounty; for ctf it self-creates the moment the next step\n(`eval_metrics.py --write`, already run once by close-out.py at SOLVED) writes to it, so it may not\nexist yet if you are running this by hand before close-out fired:\n\n**Auto first - REAL numbers, no estimation:**\n```bash\npython3 scripts\u002Feval_metrics.py $ENG --write\n```\nThis injects a `## Metrics (auto)` block from the hook telemetry (`.events.jsonl`) + the session\ntranscript recorded in `.metrics.json`: exact skill \u002F hook \u002F tool call counts, auto-detected drift\nsignals (every scope-guard block, state-discipline nudge, leak-warn), the start->finish time delta +\nidle-filtered active time, and the box's token usage (windowed to its active period). NEVER hand-type\nthese - the transcript is ground truth (a real box's true output was ~12x an earlier hand-estimate).\n\n**Then the judgement half - only you can write it:** in the same file, fill the Drift moments (what\nthe auto drift-count can't say: WHY each drift happened + the one-line fix), What went right, and the\nScores. The auto block counts drift; you narrate its cause. Dead-ends come from `Deadends.md`.\nUnlike the retro (generic, tracked), eval.md is per-engagement and may name specifics (gitignored `targets\u002F`).\n\nOnly after Phase 0 is done, proceed to the knowledge harvest below.\n\n## Steps (Phases 1-7: wiki knowledge harvest)\n\n### 1. Resolve the engagement\n```bash\nENG=$(cat targets\u002Factive.md)\nTYPE=$(grep -m1 engagement_type targets\u002F$ENG\u002Fstate.md | cut -d: -f2 | tr -d ' ')\n```\nConfirm close-out (`## STATUS:` heading in `state.md`) or that the operator asked\nexplicitly. Do not harvest an in-flight engagement unless asked.\n\n### 2. Inventory what the engagement taught\nRead the full engagement and list every candidate GENERIC lesson:\n- `state.md` - tech\u002Fservice\u002Fversion that mattered and how it was handled.\n- `loot.md` - default or vendor-known creds (NOT client-set passwords).\n- `Killchain.md` + `walkthrough.md` - the chain that actually worked, with exact commands.\n- `Deadends.md` - what failed. Negative knowledge is reusable: a bypass that does NOT\n  work on tech X, a default cred changed in vendor version Y, a false-positive pattern.\n- `Vuln-index.md` \u002F `Vulns\u002F` - findings and their reusable exploitation technique.\n- `Approach.md`, `log.md` - anything else non-obvious that recurs.\n\nA candidate is worth harvesting only if it is REUSABLE on the next engagement. Skip\none-off client trivia.\n\n### 3. Generalize + strip client specifics\nRewrite each candidate to its generic form: product + technique\u002Fcred\u002Fendpoint +\nimpact. Drop every client host, IP, domain, and client-set credential value. Client\nspecifics stay under `targets\u002F\u003Ceng>\u002F`. (The leak gate will refuse anything that slips\nthrough, but strip up front - do not lean on the gate.)\n\n**Compact to the DELTA, not a re-explanation.** A staged candidate is the MINIMAL new reusable\nprimitive + its one gotcha (a technique delta is typically \u003C ~15 lines), NOT a re-teach of what the\ntarget page already covers. Before promote, re-read the target section (Phase 4) and cut every line\nalready there -- keep ONLY what the wiki did not have. A verbose dump that repeats known material is\na failed harvest even if it promotes.\n\n### 4. Dedup against the wiki (the skip rule)\nFor each generic candidate, search the wiki first:\n```\nmcp__wiki-search__qmd_query   # semantic: concept \u002F technique \u002F intent\nmcp__wiki-search__qmd_search  # keyword: exact tool name, CVE id, payload string\n```\nFind the home page (one class = one page). Read only its frontmatter and the relevant\nsection. If the technique\u002Fpayload\u002Fcred is ALREADY covered there, SKIP it - wiki has it.\nKeep only the delta. This step is the whole point: \"extract stuff we haven't had yet\".\n\n### 5. Route each survivor through stage\nNever hand-edit `wiki\u002F` with engagement-derived content. Stage it:\n```bash\n# default \u002F vendor-known credential -> the cred cheatsheet\npython3 scripts\u002Fwiki-stage.py --kind default-cred --slug \u003Cproduct>-default \\\n  --body '| \u003Cproduct> | \u003Cversion> | \u003Cuser> | \u003Cpass> | observed | \u003Cgeneric note> |'\n\n# reusable request \u002F payload pattern -> the api-request cheatsheet\npython3 scripts\u002Fwiki-stage.py --kind api-pattern --slug \u003Cproduct>-\u003Cendpoint> \\\n  --body '| \u003Cproduct> | \u003Cendpoint> | \u003Cmethod> | \u003Crequest\u002Fpayload> | \u003Cauth> | \u003Creveals> |'\n\n# technique \u002F bypass \u002F tool-gotcha that ENRICHES an existing page\npython3 scripts\u002Fwiki-stage.py --kind technique --slug \u003Cslug> \\\n  --target-page techniques\u002F\u003Carea>\u002F\u003Cpage>.md\n# then edit targets\u002F$ENG\u002Fwiki-candidates\u002F\u003Cslug>.md to APPEND the generic '## Heading' body\n# (append below the staged frontmatter -- do NOT overwrite the file, or you drop `status: pending`\n#  and `wiki-promote --list` will show nothing to promote)\n```\n\n**Genuinely new class with no home page:** `wiki-promote` merges into an EXISTING\npage and skips a missing target. So first create a CONTENT-FREE generic scaffold\n(frontmatter per `docs\u002Fpage-types.md` + the section headings only, zero engagement\ndata), then stage the body against it so the substance still arrives through the gate:\n```bash\n# 1. scaffold wiki\u002Ftechniques\u002F\u003Carea>\u002F\u003Cslug>.md: frontmatter + empty section headings\n# 2. stage the generic body:\npython3 scripts\u002Fwiki-stage.py --kind technique --slug \u003Cslug> \\\n  --target-page techniques\u002F\u003Carea>\u002F\u003Cslug>.md\n```\nFor an external source that also informs the lesson (a CVE writeup, an advisory), hand\nthat part to `Skill(research-ingest)` rather than duplicating it here.\n\n### 6. Promote through the leak gate\n```bash\npython3 scripts\u002Fwiki-promote.py --list            # review pending candidates\npython3 scripts\u002Fwiki-promote.py --review \u003Cslug>    # read one in full\npython3 scripts\u002Fwiki-promote.py --promote all      # leak-checked merge + re-index\n```\nPromote runs `check-leaks.sh` on each body and refuses (writes nothing) on a client\nmarker. Report what promoted, what was refused, and why. `--promote` also appends a\ngeneric delta-yield tally (date, engagement_type, count) to `docs\u002Fwiki-delta-log.md`;\na zero-yield harvest is worth a sanity check (was the sweep actually run, or is the wiki\nalready saturated for this box).\n\n### 7. Re-index, lint, self-clear\n```bash\npython3 scripts\u002Flint-wiki.py -q        # must be clean (broken links, stale index)\n# only if a NEW page was created:\npython3 scripts\u002Fgen_index.py && python3 scripts\u002Fbuild_moc.py && qmd update\n# GATE: do NOT self-clear with eval.md's judgement half blank (Phase 0d). This exits 1 and\n# refuses the marker until Drift moments \u002F What went right \u002F Score are filled -- learn self-cleared\n# with the human half empty on a real box, so this is enforced, not advisory.\npython3 scripts\u002Feval_metrics.py $ENG --check-judgement || { echo \"fill eval.md judgement half first\"; exit 1; }\ntouch targets\u002F$ENG\u002F.learn-done         # marks this engagement's learn pass done\n```\nLog one GENERIC line to `session\u002Flog.md` (e.g. \"learn: promoted 3 -> jwt-attacks,\ndefault-credentials\") and an audit line to `targets\u002F$ENG\u002Flog.md`. Never put a client\nhost\u002FIP\u002Fdomain in either.\n\n## Guardrails\n- **Generic only, gate-enforced.** All engagement-derived writes go stage -> promote;\n  the leak check is the code boundary. A new page is a content-free scaffold; its\n  substance still comes through the gate.\n- **Dedup first.** Enrich the existing page; do not create a second page for a class\n  wiki already covers. `qmd` before staging, every time.\n- **No fabrication.** Harvest only what the engagement files actually record. If a\n  lesson is not backed by the state\u002Floot\u002FKillchain\u002Fwalkthrough\u002FDeadends record, drop it.\n- **Read-only on the engagement.** Phase 0 may edit HARNESS files (`skills\u002F`, `skills\u002Fhooks\u002F`,\n  `scripts\u002F`, `tests\u002F`, `docs\u002Fsuperpowers\u002Fharness-retro.md`) and Phases 1-7 write to `wiki\u002F`\n  (via the gate) and the `.learn-done` marker - but NEVER the engagement's own findings\u002Fnarrative.\n- **No offensive code in a skill body.** A tightened skill (`ctf-box`\u002F`hunt-*`) gets lesson prose +\n  `[[wikilinks]]` only. Runnable exploit\u002Frootkit\u002FPoC code lives in `wiki\u002F` (through the Phase 5-6\n  gate) and is referenced by link, never inlined - the skill loads into API context wholesale and\n  the classifier flags raw offensive code (it has). Reversing walks, syscall PoCs, magic-signal\n  dumps, CVE exploits = wiki page + a one-line skill pointer.\n- **Phase 0 changes ship green.** Any hook\u002Fscript edit lands with a test and a passing\n  `python3 -m pytest tests\u002F -q`; do not leave the harness broken by an improvement.\n- **Retro log is generic + tracked.** `harness-retro.md` describes harness failures generically\n  (no client host\u002FIP\u002Fcred); the leak boundary still applies.\n\nReport: **Phase 0** - drifts found, root cause each, harness changes made vs proposed, retro-log\npath, test status. **Phases 1-7** - candidates found, skipped-as-already-in-wiki, staged,\npromoted, refused, pages touched (and any new page created), lint status.\n",{"name":5,"description":7},"main","d69096e239f2a89aabc2cd078c47664a9d3b1e11","04-安全渗透\u002Flearn-TORCH",3,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts","https:\u002F\u002Fgithub.com\u002Flza6","个人vx：Tf00798\r\n个人✈：@TFGY999","2026-10-08 14:29:22",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add lza6\u002FArmor-piercing-system-prompts --skill learn",{"tool":38,"command":39},"bunx","bunx skills add lza6\u002FArmor-piercing-system-prompts --skill learn",{"tool":41,"command":42},"pnpm","pnpm dlx skills add lza6\u002FArmor-piercing-system-prompts --skill learn","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F04-安全渗透\u002Flearn-TORCH\u002FSKILL.md",true,[],[47,55,60,67,74,81],{"skillNo":48,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":51,"categories":52,"keywords":53,"publishedAt":54},"01M7FFG485KY83TD6HSC4YPCMC","obsidian-clipper-template-creator","Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator\u002FSKILL.md",[],[],"2026-11-16 10:05:34",{"skillNo":56,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":57,"categories":58,"keywords":59,"publishedAt":54},"01M7FFG0T6AZ0STEJPSB3CX84K","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator-antigravity-awesome-skills-main\u002FSKILL.md",[],[],{"skillNo":61,"name":62,"nameZh":6,"description":63,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":54},"01M7FFFY0QEM14SKWPDYGMXM2P","python-api","Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP\u002FmAP semantics, and source-style import diagnostics.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Fpython-api\u002FSKILL.md",[],[],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":54},"01M7FFFVZQZ2GJ0C0TGP7JHD20","file-evaluation","Folder-based PASCAL VOC AP\u002FmAP evaluation from ground-truth and detection text files.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Ffile-evaluation\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":54},"01M7FFFSFNYSE77SJ6PE1CK2DA","object-detection-metrics","Guides legacy Object-Detection-Metrics PASCAL VOC AP\u002FmAP evaluation from detection text folders and direct Python API classes.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":54},"01M7FFFP0GCK1RVXBCPQ8APBMT","nx-workspace-patterns","Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fnx-workspace-patterns\u002FSKILL.md",[],[]]