Use Cases

Use-Cases

1. Cold agent resumes project

Scenario: Return to project after 2 weeks. Agent has no chat history.

saipen continue

Agent reads STATE.md (phase: BUILD, task: T-7), BOARD.md (T-7 in DOING), LOG tail. Executes next_action: PHASE BUILD T-7. Resumes coding without asking "what was I doing?"

2. Overnight autonomous evolution

Scenario: Leave agent running with goal mode.

saipen goal 'make the UI responsive'

Agent: PLAN -> 5 tickets. Builds, verifies, reviews, ships. Then HUNT -> clean -> ADD -> finds symmetry gaps -> tickets -> builds -> ships. Hits wave 3 cap. Writes WAIT: safety valve. Stops. Morning: 3 releases, 15 tickets, zero questions asked.

3. Parallel subSaipen research

Scenario: Need to audit codebase while building a feature.

saipen sub spawn saihunt saipen sub spawn saipython

saihunt scans 6 HUNT categories. saipython finds Python issues (untyped defs, bare excepts), patches them in its own kitchen/pen/ sandbox. Both write OUTBOXes. Main agent collects after current BUILD.

Collect discipline: main agent runs git status against whole tree before folding. A confused sub that wrote outside its folder gets caught by boundary check (v7.98.0+).

4. Safety valve prevents runaway

Scenario: Agent running autonomously for hours. Goal caps prevent infinite loops.

After 3 waves or 20 tickets: agent stops, writes WAIT: safety valve -- X waves completed, re-invoke saipen goal to authorize more. User reviews, re-authorizes or stops. Never burns unlimited compute.

5. Multi-vendor handoff

Scenario: Start with Claude, continue with Gemini, ship with Codex.

Claude: saipen goal 'build auth' -> builds, ships v0.5.0. Gemini: saipen continue -> reads STATE, HUNTs, ADDs next feature. Codex: saipen ship -> bumps to v0.6.0, tags, pushes.

All three agents from the same STATE.md. No context loss between vendors.

6. Blocked project recovery

Scenario: T-14 depends on T-12 which is stuck.

Agent at DONE: empty board? No, T-14 blocked.
BLOCKED: T-14 needs T-12, T-12 needs human decision.
WAIT: blocked -- T-12 needs UX mockup approval before build can start

Human sees WAIT, provides mockup. Next continue: T-12 unblocked, builds, ships, T-14 builds on top.

7. First-publish gate

Scenario: First release of a new project.

SHIP detects no origin, writes: WAIT: first-publish -- confirm repo name 'my-project' and public/private before I push

Human confirms. Agent pushes, tags, creates first release. One-way door protected.

8. Portable floor validation

Scenario: CI server without Python.

GitHub Actions (has Python) -> runs tools/validate.py. Minimal Docker container (no Python) -> runs tests/validate.sh or tests/validate.ps1.

Both probe all 9 STATE fields, catch read-only bans, check WAIT categories. Proved red: tools/audit_floor.py breaks a scratch project 20 ways, asserts both halves report each failure (v7.100.0+).

9. Translation pipeline

Scenario: Locale packs for 32 languages.

ee

Agent quarantines in .saipen/saitranslate/, scans both real docs and real in-app UI strings, builds all 32 locales plus дед voice, verifies coverage, and writes a complete status: ready OUTBOX. It never touches the main tree or pushes. Later eee accepts only that fresh ready package, integrates its declared payload through VERIFY→REVIEW→SHIP, and pushes; anything incomplete replies Not ready: run ee first. with zero main write.

10. Crash recovery

Scenario: Agent crashes mid-BUILD. STATE says phase: BUILD but work incomplete.

Next agent: RECOVER. STATE stale? BOARD shows T-17 in DOING with old claim_time. LOG says RUN: BUILD T-17 with no VERIFY. Adopts T-17 (new claim_time). Continues BUILD. No data loss, no duplicate work.

11. Citation resolution (v7.101.0+)

Scenario: A shipped doc references SS N.N that was renumbered. Nothing signals the break.

SAIPEN's validator now resolves every SS N.N and phases/.md reference across 103 shipped documents. Dead citation = FAIL. Caught: 7 orphaned citations when SS 1.11 was renumbered, phases/add.md teaching a RETURN SCOUT rule phases/hunt.md had already replaced, KNOWLEDGE/traps.md describing LOG drift check as WARN when it is FAIL.

12. Push-claim honesty (v7.99.0+)

Scenario: Agent claims "shipped, committed, pushed" in next_action. Three CI commits never left the machine. Nothing contradicts because nothing looks.

Now FAILs: STATE.next_action mentioning push while commits sit local-only. Degrades to WARN only when no git or no upstream. The repository itself adjudicates the claim.

13. Continue from a subfolder (v7.132.0+)

Scenario: You cd into src/auth/, type saipen continue. Where does memory land?

Before v7.132.0: checkpoints were relative to ambient cwd — a cd could make STATE.md name a different project without any error. Now BOOT binds one project root for the session: Git worktree/common-dir identity, linked worktrees use the main worktree's memory, non-Git projects use the nearest ancestor with .saipen/. An unowned cwd fails instead of creating a second memory tree. From another machine: saipen continue --project-root <path> overrides deliberately.

14. Release guard vs stray tags (v7.130.0+)

Scenario: A leftover annotated tag sits in local state with the same name as a release. git push --follow-tags publishes it beside the intended tag.

SHIP now pushes the branch and one exact refs/tags/vVERSION in separate commands. --tags/--follow-tags are forbidden for releases; the temporary bare-origin probe proved the prescribed refspec publishes only the intended tag. Same intent, one more line: "publish the release you named."

15. Shortcuts on a Russian layout (v7.152.0+)

Scenario: You type on a Russian keyboard layout. сс (two Cyrillic 'с') looks like cc but is not the same codepoint.

The recognizer normalizes through the lowercase confusable set (аеорсух) before matching — six shortcuts have homoglyph twins (сс, ссс, аа, ее, еее, рр), seven have none. Replying "unrecognized command" to a key the user pressed correctly was the worst available response, and it landed on the most-used shortcut. The table stays Latin: canonical spelling, not the only accepted form.

16. The brake vs the status report (v7.149.0+)

Scenario: You want to stop a run, but ss used to mean status. s alone collides across stop/set/ship/status/scout.

ss = saipen stop, sss = saipen status; cc = saipen goal, ccc = continue plus shipping. Length has no global meaning: sss is read-only while eee/qqq are ready-only integration chains. Read the exact row; never invent an undeclared repeated form.