condense-memory

<!-- cloud_safe: noop -->

Experimental. Output-side telegraph dialect did not meet the kill-criterion in internal/bench/reports/telegraph-v1.md (vs_terse median −9.27 %). Input-side memory condensation is an orthogonal use case: the savings target the always-loaded memory budget, not the reply stream. Treat ship-criterion as per-target measurement, not the v1 verdict.

When to use

Use when:

• An always-loaded memory file (AGENTS.md, CLAUDE.md, .cursorrules, GEMINI.md, .windsurfrules) is close to or above the host tool's char budget and the maintainer wants to recover input-token headroom.
• A consumer-shipped templates/AGENTS.md is failing the agents-md-thin-root cap and the pointer-extraction options are exhausted.
• The maintainer asks to "condense this memory file" or "shrink AGENTS.md" or names input-side telegraph.

Do NOT

• Condense a reply, commit message, PR body, ticket summary, or any deliverable written for a human reader — those are carve-outs in telegraph-speak § Carve-outs and stay verbatim.
• Condense a path matching the sensitive-file denylist (.env*, .netrc, credentials*, secrets*, id_rsa*, *.pem|key|p12|pfx|crt|cer|jks, .ssh/*) — the script refuses with SensitivePathError and so should you.
• Condense a generated file (dist/agent-src/, .augment/, .claude/, .cursor/, .clinerules/, .windsurfrules) — edit the source in src/ and regenerate via the package's sync + generate-tools scripts (scripts/condense.sh --sync + scripts/condense.ts --generate-tools).
• Hand-edit a condensed memory file in place — run --decondense first; the next condense pass refuses on body-hash drift (CondensationRefused).
• Commit the condensed file without committing the matching .original.md backup — round-trip breaks otherwise.

Procedure

1. Analyse the target first. Before any write, inspect the target with view or wc -l to confirm it is an always-loaded memory file (AGENTS.md, CLAUDE.md, .cursorrules, GEMINI.md, .windsurfrules), is not generated, and has prose paragraphs to condense (a pointer-only Thin-Root file may net near-zero). Skip the rest of the procedure if any check fails.
2. Check denylist gate. Run ./scripts-run src/scripts/condense_memory <path> --check — exit 0 = safe; exit 2 = denylist hit, stop and surface the refusal.
3. Record baseline. wc -c <path> — capture pre-condensation char count for the commit message.
4. Condense. ./scripts-run src/scripts/condense_memory <path>. The script writes <path>.original.md (verbatim backup) and rewrites <path> with original_sha256: + condensed_at: frontmatter.
5. Inspect the diff. Eyeball every Iron-Law fence, numbered-options block, code fence, backtick span, ❌/⚠️/✅ line, and frontmatter pair — all must be byte-identical. Body prose may have lost articles (the/a/an) and auxiliaries (is/are/was/be/that/which).
6. Validate idempotency. Re-run ./scripts-run src/scripts/condense_memory <path> — clean re-run is a no-op (body hash matches). Non-zero exit = stop, escalate.
7. Commit both files together. <path> and <path>.original.md ship as a pair. The backup is the rollback path; never commit one without the other.
8. Rollback path. If readability fails review at step 5: ./scripts-run src/scripts/condense_memory <path> --decondense restores the backup and deletes .original.md.

Output format

The maintainer-facing report after invoking the script MUST contain, in this order:

1. Diff line — pre/post wc -c as a single line (AGENTS.md: 2,891 → 2,453 chars (−15.1 %)).
2. Backup path — full path of the .original.md backup so the maintainer can verify it landed on disk.
3. Carve-out check — one line confirming the seven carve-out classes round-tripped (carve-outs: 7 classes preserved · idempotent re-run: clean).
4. Exit-code surface — on failure, surface the verbatim exit code and exception name (SensitivePathError → exit 2, CondensationRefused → exit 3, FileNotFoundError → exit 4); do not paraphrase.

Do not narrate the algorithm, the grammar rules, or the carve-out theory — the rule and this skill document the contract; the output reports the result.

Carve-outs — byte-for-byte preserved

Mirrors the seven carve-out classes in telegraph-speak. The condensation engine in src/scripts/condense_memory.ts preserves:

1. Triple-backtick fences — any language, any depth.
2. Numbered-options lines — ^>?\s*\d+\.\s plus the **Recommendation:** / **Empfehlung:** label.
3. Backtick spans — file paths, command names, identifiers inside body prose.
4. Status / error markers — lines starting with ❌, ⚠️, ✅.
5. Iron-Law ALL-CAPS lines — ^[A-Z][A-Z0-9 ,.\-_/']{3,}$.
6. Frontmatter blocks — --- fence pairs at the head of the file.
7. Mode markers per role-mode-adherence.

Mangling any of these breaks the Iron-Law surface the host tool reads. The unit tests in tests/scripts/condense_memory.test.ts lock each carve-out class as a regression case.

Idempotency contract — Step 9 guard

The script is idempotent on clean re-runs: running it twice on the same target is a no-op because the body hash matches the recondensed hash. The script refuses on body drift:

State	Outcome
No frontmatter SHA marker	Condense + write backup + inject SHA.
SHA marker present, body re-condenses to same hash	No-op (return target unchanged).
SHA marker present, body hash diverged	Refuse with CondensationRefused exit 3.

If you need to edit a condensed memory file, run --decondense first, edit the restored .original.md content, then re-run the condenseor. Never hand-edit the condensed body — the next CI run will either silently corrupt your edit (if it happens to re-condense to the same shape) or hard-fail the next condense pass.

Sensitive-path gate

Every read path passes through src/scripts/validate_safe_paths.ts assert_safe() before bytes leave disk. The gate is the security floor for Phase 2 (input-side condensation) per step-16-telegraph-substance.md Phase 0; rollback of the gate is rollback of this skill.

CLI exit codes:

• 0 — condense / decondense / check succeeded.
• 2 — SensitivePathError (path matched denylist).
• 3 — CondensationRefused (body hash diverged from frontmatter SHA).
• 4 — FileNotFoundError (no .original.md backup to restore).

Gotchas

• Body-hash drift after manual edit — hand-editing the condensed body breaks the original_sha256: invariant. The next condense pass refuses with CondensationRefused (exit 3). Recovery: --decondense, edit the restored body, re-condense.
• .original.md backup missing on --decondense — exit 4 (FileNotFoundError). Either someone deleted the backup or --decondense already ran. Restore from git history; never regenerate the backup by hand (the regenerated content would not be byte-identical).
• Denylist false positive — a sensitive-looking filename outside the denylist surface (project-specific naming) will still pass assert_safe(). The denylist is necessary but not sufficient; the maintainer is responsible for never feeding secrets to the condenseor.
• Frontmatter ordering with existing keys — if the target already has frontmatter, the condenseor preserves existing keys, drops any prior original_sha256: / condensed_at: entries, and appends the new pair. Other agents reading the file should treat the SHA + timestamp pair as the canonical condensation marker, not the file size.
• Negative savings on pointer-heavy files — a templates/AGENTS.md that already follows Thin-Root (≥ 40 % pointers, ≥ 60-char why-clauses) has little prose left to drop; condensation may net near-zero or even add bytes via frontmatter. Run agents-md-thin-root first to maximise pointer share, then measure whether this skill still pays.
• Generated-tree drift — condensing src/agent-src/templates/AGENTS.md does NOT propagate to .augment/, .claude/, etc. until the package's sync + generate-tools scripts run (scripts/condense.sh --sync + scripts/condense.ts --generate-tools). Always regenerate after condensing a templated file.

Measurement — when to condense

There is no published telegraph-v2 baseline for input-side savings yet (Step 11 of step-16-telegraph-substance.md ships that). Until then, the maintainer judges per-target whether the condensation pays its readability cost. Suggested workflow:

1. wc -c <path> before — record baseline char count.
2. ./scripts-run src/scripts/condense_memory <path> — condense + back up.
3. wc -c <path> after — record post-condensation char count.
4. Eyeball the diff: does the prose stay legible? Are all Iron-Law fences intact?
5. If yes → commit both <path> and <path>.original.md. If no → --decondense.

A future telegraph-v2.md will tabulate the realised input-token saving against the agents-md-thin-root 40 % pointer-ratio constraint so the maintainer has a numerical floor.

Cross-references

• telegraph-speak — runtime rule the script mirrors for input-side targets; that output-side rule does not gate this script (input-side runs regardless).
• src/scripts/validate_safe_paths.ts — Phase 0 gate; ported from upstream Telegraph 63a91ec.
• src/scripts/condense_memory.ts — implementation.
• tests/scripts/condense_memory.test.ts — regression locks for each carve-out + idempotency + denylist.
• docs/contracts/condensation-default-kill-criterion.md — v1 verdict (output-side; informs but does not gate this skill).
• agents-md-thin-root — caps the consumer-shipped templates/AGENTS.md; this skill is one tool to land under the cap.