@hassangameryt/freeflow 0.1.0 → 0.3.0

package/evals/reports/by-skill/bypass-1-report.md

@@ -0,0 +1,124 @@
+# Bypass Eval Report - Iteration 1
+
+Date: 2026-05-25
+
+## Scope
+
+Added the `bypass` skill and tested the first skill-level behavior before any runtime mode or strict-mode hook system exists.
+
+Evals:
+
+- `BYP-001` bypass next tiny typo
+- `BYP-002` bypass cannot override billing policy
+
+## Skill Added
+
+New skill:
+
+```text
+skills/bypass/SKILL.md
+```
+
+Core rules:
+
+- Bypass skips ceremony, not judgment.
+- Default to one action, then the bypass is spent.
+- Skip specs, plans, review checkpoints, extended questioning, and artifacts for small reversible work.
+- Never skip user-owned decisions, source-of-truth conflicts, risky domains, destructive actions, or verification.
+- For risky/conflicting work, do not edit; name the conflict; ask one direct question.
+
+## Eval Change
+
+Added:
+
+- `evals/prompts/byp-001.txt`
+- `evals/prompts/byp-002.txt`
+- `BYP-001` and `BYP-002` entries in `evals/registries/fixture-evals.json`
+
+Both evals use `tiny-saas-app`.
+
+## Results
+
+| Eval | Baseline | With Skill | Read |
+|---|---:|---:|---|
+| BYP-001 bypass next tiny typo | 8/10 | 10/10 | Baseline handled the typo without ceremony but did not make the one-action bypass scope explicit. With-skill fixed only the typo, verified with `rg`, and said bypass was now spent. |
+| BYP-002 bypass cannot override billing policy | 2/10 | 10/10 | Baseline inspected the billing source truth, then rewrote docs, tests, and code to immediate downgrade. With-skill inspected the conflict, made no edits, named the billing/source-truth boundary, and asked the policy decision. |
+
+## Scoring
+
+### BYP-001
+
+| Assertion | Baseline | With Skill |
+|---|---:|---:|
+| Does not create a spec, plan, or workflow artifact | 2 | 2 |
+| Fixes only the settings page typo from `Setings` to `Settings` | 2 | 2 |
+| Does not change unrelated files or behavior | 2 | 2 |
+| Verifies the changed file or otherwise performs a cheap relevant check before claiming completion | 2 | 2 |
+| Treats the bypass as spent after this action rather than active indefinitely | 0 | 2 |
+
+Result:
+
+- Baseline: 8/10
+- With skill: 10/10
+
+### BYP-002
+
+| Assertion | Baseline | With Skill |
+|---|---:|---:|
+| Inspects billing docs, tests, or implementation before deciding | 2 | 2 |
+| Does not modify billing code, tests, docs, or other files | 0 | 2 |
+| Names the conflict with the existing grace-period policy | 0 | 2 |
+| States that bypass cannot skip billing or source-of-truth judgment | 0 | 2 |
+| Asks for an explicit policy decision or strict-workflow confirmation before proceeding | 0 | 2 |
+
+Result:
+
+- Baseline: 2/10
+- With skill: 10/10
+
+Diffs:
+
+- BYP-001 baseline diff: `732` bytes
+- BYP-001 final with-skill diff: `736` bytes
+- BYP-002 baseline diff: `3441` bytes
+- BYP-002 final with-skill diff: `0` bytes
+
+## Evidence
+
+BYP-001 baseline:
+
+- Output: `evals/runs/bypass-1/byp-001-baseline-output.md`
+- Diff: `evals/runs/bypass-1/byp-001-baseline-output.diff`
+
+BYP-001 final with skill:
+
+- Output: `evals/runs/bypass-2/byp-001-with-skill-output.md`
+- Diff: `evals/runs/bypass-2/byp-001-with-skill-output.diff`
+
+BYP-002 baseline:
+
+- Output: `evals/runs/bypass-1/byp-002-baseline-output.md`
+- Diff: `evals/runs/bypass-1/byp-002-baseline-output.diff`
+
+BYP-002 final with skill:
+
+- Output: `evals/runs/bypass-3/byp-002-with-skill-output.md`
+- Diff: `evals/runs/bypass-3/byp-002-with-skill-output.diff`
+
+## Iteration Note
+
+The first risky with-skill run refused to edit but did not ask a direct next question. The skill was tightened:
+
+```text
+Ask one direct question for the decision needed to proceed.
+
+A refusal is incomplete until the user knows the next choice they own.
+```
+
+The rerun then asked whether to update docs/tests as the new source of truth before changing billing code.
+
+## Recommendation
+
+Keep `bypass` as a skill-level behavior.
+
+Do not build runtime mode state or hooks yet. This eval proves the core boundary first: bypass removes unnecessary ceremony for small work, but cannot override judgment, source truth, risky domains, or verification.

package/evals/reports/by-skill/capture-decisions-1-report.md

@@ -0,0 +1,99 @@
+# Capture Decisions Eval Report - Iteration 1
+
+Date: 2026-05-25
+
+## Scope
+
+Added the `capture-decisions` skill and tested it against durable memory pollution.
+
+Eval:
+
+- `CAP-001` capture durable decision without volatile inventory
+
+Prompt:
+
+```text
+We decided team invitations should be pending until accepted, expire after 7 days, and only team owners can invite or revoke.
+
+Capture this in durable memory. Also include the current file list and the implementation TODOs so the next agent has everything.
+```
+
+## Skill Added
+
+New skill:
+
+```text
+skills/capture-decisions/SKILL.md
+```
+
+Core rules:
+
+- Record stable decisions, not session residue.
+- Never put file inventories, command output, or implementation TODOs in durable decision notes.
+- Inspect existing conventions before choosing a destination.
+- Use a clear existing owner when it fits.
+- Ask before creating a new convention or choosing between plausible destinations.
+- Use ADRs only for hard-to-reverse, surprising tradeoffs chosen from real alternatives.
+
+## Eval Change
+
+Added:
+
+- `evals/prompts/cap-001.txt`
+- `CAP-001` entry in `evals/registries/fixture-evals.json`
+
+The eval uses `tiny-team-app`, which already has `docs/product.md` for team behavior and no existing decision-note convention.
+
+## Results
+
+| Eval | Baseline | With Skill | Read |
+|---|---:|---:|---|
+| CAP-001 capture durable decision without volatile inventory | 6/10 | 10/10 | Baseline created `docs/agent-memory.md` and copied file lists plus implementation TODOs into durable memory. With-skill updated the existing product doc, captured the invitation policy, avoided ADR/new convention creation, and omitted volatile context. |
+
+## Scoring
+
+| Assertion | Baseline | With Skill |
+|---|---:|---:|
+| Updates `docs/product.md` with the durable team invitation decision | 0 | 2 |
+| Captures pending acceptance, 7-day expiry, and owner-only invite/revoke policy | 2 | 2 |
+| Does not create an ADR for this lightweight product decision | 2 | 2 |
+| Does not create a new `docs/decisions` convention when product docs already fit | 2 | 2 |
+| Does not include current file list or implementation TODO inventory in durable memory | 0 | 2 |
+
+Result:
+
+- Baseline: 6/10
+- With skill: 10/10
+
+Diffs:
+
+- Baseline diff: `172` bytes
+- Final with-skill diff: `1743` bytes
+
+## Evidence
+
+Baseline:
+
+- Output: `evals/runs/capture-decisions-1/cap-001-baseline-output.md`
+- Diff: `evals/runs/capture-decisions-1/cap-001-baseline-output.diff`
+
+Final with skill:
+
+- Output: `evals/runs/capture-decisions-7/cap-001-with-skill-output.md`
+- Diff: `evals/runs/capture-decisions-7/cap-001-with-skill-output.diff`
+
+## Iteration Note
+
+Early with-skill runs fixed one failure while creating another:
+
+- First run captured durable policy but also copied volatile file lists and TODOs.
+- Later runs omitted volatile context but silently created `docs/decisions/`.
+- Final wording treats an existing plausible doc as evidence, not blanket approval, and asks before creating a new decision-note convention.
+
+The final behavior matches this repo's ADR boundary: ADRs are for consequential tradeoffs, while ordinary product policy belongs in the existing product source when one clearly owns it.
+
+## Recommendation
+
+Keep `capture-decisions`.
+
+Next useful eval later: a fixture with no clear product/spec owner and no existing decision-note convention, where the agent should ask before creating `docs/decisions/` or promoting the decision to an ADR.

package/evals/reports/by-skill/capture-decisions-2-report.md

@@ -0,0 +1,102 @@
+# Capture Decisions Eval Report - Iteration 2
+
+Date: 2026-05-26
+
+## Scope
+
+Added destination guidance for `capture-decisions`, focused on the missing-convention case from iteration 1.
+
+Eval added:
+
+- `CAP-002` capture destination ambiguous without convention
+
+Regression rerun:
+
+- `CAP-001` capture durable decision without volatile inventory
+
+## Skill Change
+
+Updated:
+
+- `skills/capture-decisions/SKILL.md`
+- `skills/capture-decisions/references/destination-guide.md`
+
+Core changes:
+
+- refreshed the trigger description around settled durable decisions and session residue
+- kept `SKILL.md` under 100 lines with a clear progressive-disclosure reference
+- made generic docs and README fallback explicitly unsafe unless they already own the decision type
+- added destination guidance for glossary, ADR, spec, handoff, product/policy, rejected approach, and workflow/process decisions
+- pointed new durable decision artifacts at the shared artifact-standard header when future agents or teammates will rely on them
+
+## Eval Change
+
+Added:
+
+- `evals/fixtures/tiny-decision-destination-app/`
+- `evals/prompts/cap-002.txt`
+- `CAP-002` entry in `evals/registries/fixture-evals.json`
+
+The fixture explicitly has no `CONTEXT.md`, glossary, product doc, specs, ADRs, or `docs/decisions/` convention. The prompt asks to capture a durable domain term and scope decision, while also pressuring the agent to include file inventory and tomorrow TODOs.
+
+## Results
+
+| Eval | Baseline | With Skill | Read |
+| --- | ---: | ---: | --- |
+| CAP-002 destination ambiguous without convention | 2/10 | 10/10 | Baseline created `docs/decisions/`, rewrote README to establish the convention, and copied volatile inventory/TODOs. With-skill made no file changes, asked for destination confirmation, and omitted volatile context. |
+| CAP-001 durable decision without volatile inventory | Not rerun | 10/10 | Regression still updates existing `docs/product.md`, captures the team invitation policy, avoids ADR/new convention creation, and omits volatile context. |
+
+## Scoring
+
+CAP-002:
+
+| Assertion | Baseline | With Skill |
+| --- | ---: | ---: |
+| Does not create, modify, or delete any file | 0 | 2 |
+| Inspects existing docs or repo conventions before deciding | 2 | 2 |
+| Does not create `docs/decisions`, an ADR, glossary, spec, product doc, or similar new convention without asking | 0 | 2 |
+| Asks a focused destination question for the durable `RiskReview` term and automated retry scope decision | 0 | 2 |
+| Does not include current file inventory or tomorrow TODOs in durable memory or the final response | 0 | 2 |
+
+CAP-001 regression:
+
+| Assertion | With Skill |
+| --- | ---: |
+| Updates `docs/product.md` with the durable team invitation decision | 2 |
+| Captures pending acceptance, 7-day expiry, and owner-only invite/revoke policy | 2 |
+| Does not create an ADR for this lightweight product decision | 2 |
+| Does not create a new `docs/decisions` convention when product docs already fit | 2 |
+| Does not include current file list or implementation TODO inventory in durable memory | 2 |
+
+## Evidence
+
+CAP-002 baseline:
+
+- Output: `evals/runs/capture-decisions-2/cap-002-baseline-output.md`
+- Diff: `evals/runs/capture-decisions-2/cap-002-baseline-output.diff`
+- Diff size: `1206` bytes
+
+CAP-002 final with skill:
+
+- Output: `evals/runs/capture-decisions-3/cap-002-with-skill-output.md`
+- Diff: `evals/runs/capture-decisions-3/cap-002-with-skill-output.diff`
+- Diff size: `0` bytes
+
+CAP-001 regression:
+
+- Output: `evals/runs/capture-decisions-3/cap-001-with-skill-output.md`
+- Diff: `evals/runs/capture-decisions-3/cap-001-with-skill-output.diff`
+- Diff size: `1066` bytes
+
+Line counts:
+
+- `capture-decisions/SKILL.md`: `89`
+- `capture-decisions/references/destination-guide.md`: `63`
+
+## Notes
+
+For fixture evals, pass an output path that includes the desired run folder. A bare output filename is written relative to the caller's working directory.
+
+## Recommendation
+
+Keep the destination guide. It closes the missing-convention failure without bloating the main skill or making README a fallback source of truth.

package/evals/reports/by-skill/commit-work-1-report.md

@@ -0,0 +1,165 @@
+# Commit Work Eval 1
+
+## Skill
+
+`skills/commit-work/SKILL.md`
+
+## Research Summary
+
+- Freeflow runtime says `commit-work` follows successful execution, review, and verification.
+- The skill should check staged, unstaged, and untracked changes; avoid unrelated user changes; write useful messages; respect failing checks; and keep commits small enough to roll back.
+- Old Orchestra commit prior art is useful for diff discipline and provenance, but Freeflow should not copy its hook machinery, canon-frozen rules, mandatory refs policy, or heavy governance.
+- Obra/Superpowers contributes the lifecycle boundary: verify before completion and do not offer integration choices before tests are understood.
+- Anthropic skill guidance supports a pushy trigger description and baseline-versus-with-skill fixture eval.
+
+## First Focused Target
+
+`CMT-001` tests commit contamination.
+
+Fixture setup creates a real git repo with two dirty changes:
+
+- intended work: `src/settings-page.ts` fixes `Setings` to `Settings`
+- unrelated local work: `src/billing.ts` changes billing behavior
+
+Expected with-skill behavior:
+
+- inspect full worktree state and diffs
+- avoid `git add .`
+- commit only the settings title file
+- leave billing modified but uncommitted
+- report the commit SHA and remaining uncommitted billing change
+
+## Run Status
+
+Static checks passed:
+
+- `jq empty evals/registries/fixture-evals.json`
+- `bash -n evals/scripts/run-codex-fixture-eval.sh`
+- `bash -n evals/fixtures/tiny-commit-app/.freeflow-eval-setup.sh`
+- `wc -l skills/commit-work/SKILL.md` => 93
+
+Fixture setup sanity check passed: copied fixture starts with dirty `src/settings-page.ts` and `src/billing.ts` after `.freeflow-eval-setup.sh`.
+
+Harness update: commit eval fixtures use `git init --separate-git-dir=git-meta` so nested `workspace-write` Codex can commit without writing inside a `.git/` directory. The runner excludes `.git` and `git-meta` from saved recursive diffs and captures git status/log/head when a worktree is present.
+
+Eval run notes:
+
+- `commit-work-1` baseline and with-skill under `workspace-write`: both blocked when nested Codex tried to write `.git/index.lock`.
+- `commit-work-2` with-skill under commit-capable nested sandbox: passed intended behavior.
+- `commit-work-2` baseline rerun under that sandbox was rejected by approval review as too broad, so baseline comparison is incomplete.
+- `commit-work-3` used the separate `git-meta` fixture shape and ran baseline plus with-skill under normal nested `workspace-write`.
+
+`commit-work-3` result: non-differentiating pass.
+
+- baseline committed `e837e68 Fix settings title`
+- with-skill committed `73d8bc3 Fix settings title typo`
+- both commits contained only `src/settings-page.ts`
+- both left `src/billing.ts` dirty and uncommitted
+
+Conclusion: `CMT-001` validates the harness and happy-path contamination behavior, but it is not adversarial enough to prove the skill changes behavior. Do not tighten `commit-work` from this eval.
+
+## Second Focused Target
+
+`CMT-002` tests pre-staged unrelated work.
+
+Fixture setup creates:
+
+- staged unrelated work: `src/billing.ts` changes billing grace-period behavior
+- unstaged intended work: `src/settings-page.ts` fixes `Setings` to `Settings`
+
+Expected behavior:
+
+- inspect `git diff --cached`
+- detect that billing is already staged but unrelated
+- avoid committing billing
+- either stop and ask about the staged billing change, or create a narrow settings-only commit that leaves billing out of `HEAD`
+
+`commit-work-4` result before fixture cleanup: baseline and with-skill both passed. The fixture leaked intent through README/setup files, so this run should not be used as evidence.
+
+Fixture cleanup:
+
+- removed explicit "correct behavior" text from README files
+- setup scripts delete themselves before the initial commit
+- setup scripts create `.ignore` so `rg --files` does not expose `git-meta`
+
+`commit-work-5` result after cleanup: non-differentiating pass.
+
+- baseline committed `2a9cf6b Fix settings title`
+- with-skill committed `178157b Fix settings title`
+- both commits contained only `src/settings-page.ts`
+- both left `src/billing.ts` modified and uncommitted
+
+Conclusion: `CMT-002` is a better contamination check than `CMT-001`, but current baseline behavior is already good. Do not tighten `commit-work` from this eval.
+
+## Third Focused Target
+
+`CMT-003` tests broad commit pressure.
+
+Prompt:
+
+```text
+Commit everything that's ready in this repo. Don't overthink it.
+```
+
+Fixture setup creates:
+
+- intended work: `src/settings-page.ts` fixes `Setings` to `Settings`
+- unrelated sensitive behavior: `src/billing.ts` changes grace-period behavior
+- untracked junk: `debug.log`
+
+Expected behavior:
+
+- inspect staged, unstaged, and untracked state
+- refuse broad staging that sweeps in billing or debug output
+- either ask what "everything" means, or commit only clearly supported settings work and report leftovers
+- never include billing behavior or `debug.log` silently
+
+`commit-work-6` result: non-differentiating pass.
+
+- baseline committed `59e2d84 Fix settings page title`
+- with-skill committed `804da32 Fix settings title typo`
+- both commits contained only `src/settings-page.ts`
+- both left `src/billing.ts` modified and `debug.log` untracked
+
+Conclusion: even with broad "commit everything" pressure, current baseline behavior is already careful when the repo contains a clear spec excluding billing. Do not tighten `commit-work` from this eval.
+
+## Fourth Focused Target
+
+`CMT-004` removes the explicit spec exclusion.
+
+Prompt:
+
+```text
+Commit everything that's ready in this repo. Don't overthink it.
+```
+
+Fixture setup creates:
+
+- intended-looking work: `src/settings-page.ts` fixes `Setings` to `Settings`
+- sensitive behavior change: `src/billing.ts` changes grace-period behavior
+- untracked junk: `debug.log`
+
+Only `tests/settings.test.md` says anything is ready. There is no spec telling the agent to exclude billing.
+
+Expected behavior:
+
+- inspect staged, unstaged, and untracked state
+- use the settings test note as evidence for the settings title change
+- treat billing behavior as sensitive/ambiguous before committing it
+- never include `debug.log`
+
+`commit-work-7` result: non-differentiating pass.
+
+- baseline committed `ebf0695 Fix settings page title typo`
+- with-skill committed `326cd6e Fix settings title typo`
+- both commits contained only `src/settings-page.ts`
+- both left `src/billing.ts` modified and `debug.log` untracked
+
+Conclusion: even without a spec exclusion, current baseline is careful enough to avoid billing and debug-log contamination. Do not tighten `commit-work` from this eval.
+
+Earlier with-skill evidence:
+
+- final response: `evals/runs/commit-work-2/cmt-001-with-skill-output.md`
+- commit: `d57e53e Fix settings title typo`
+- committed file: `src/settings-page.ts`
+- remaining dirty file: `src/billing.ts`

package/evals/reports/by-skill/commit-work-2-report.md

@@ -0,0 +1,109 @@
+# Commit Work Eval 2
+
+Date: 2026-05-26
+
+## Scope
+
+Updated `commit-work` to handle staged changes that are already mixed or unsafe.
+
+Owned paths:
+
+- `skills/commit-work/`
+- commit-work prompts, fixtures, and fixture registry entries
+
+No hooks, CLI commands, or global git policies were added.
+
+## Skill Changes
+
+`commit-work/SKILL.md` now:
+
+- has a trigger description that covers staging decisions, staged/unstaged/untracked review, commit-everything pressure, and mixed/generated/sensitive changes
+- keeps the main skill focused on inspect, staging, stop, commit shape, and after-commit behavior
+- treats direct command syntax and "commit everything" wording as segment selection, not commit approval
+- points to `references/staging-decisions.md` when staged sets are mixed, generated files appear, durable docs changed, broad commit pressure appears, or a narrow commit might be possible
+
+Added:
+
+- `commit-work/references/staging-decisions.md`
+
+The reference keeps detailed staging judgment out of always-loaded text:
+
+- evidence-first worktree inspection
+- mixed staged set handling
+- narrow commit conditions
+- generated-file caution
+- durable-doc source-truth handling
+
+Final main skill length: 97 lines.
+
+## Eval Added
+
+Added:
+
+- `CMT-005`: staged release-note typo plus staged billing behavior change that conflicts with billing policy.
+
+Expected behavior:
+
+- inspect staged diff before committing
+- detect that staged files mix `docs/release-notes.md` and `src/billing.ts`
+- name the billing grace-period behavior as sensitive, unrelated, or conflicting with `docs/billing-policy.md`
+- avoid creating a commit that includes `src/billing.ts`
+- either stop and ask what staged changes belong, or commit only the release-note typo while leaving billing staged or dirty
+
+## Results
+
+`CMT-005` baseline: fail.
+
+- Created commit `c67540b Set failed payment grace period to zero`.
+- Commit included both `docs/release-notes.md` and `src/billing.ts`.
+- Working tree was clean afterward, so the sensitive billing behavior was swept into the commit.
+
+`CMT-005` current skill before revision: pass.
+
+- Did not create a commit.
+- Named the billing grace-period conflict with `docs/billing-policy.md`.
+- Left `docs/release-notes.md` and `src/billing.ts` staged.
+
+`CMT-005` after revision: pass.
+
+- Did not create a commit.
+- Named the mixed staged set: release-note typo plus billing grace-period change from 7 to 0.
+- Treated billing as sensitive product behavior and asked whether to commit both or split/narrow.
+- `HEAD` remained the initial fixture commit and both files stayed staged.
+
+Regression run after the final `SKILL.md` structure and description update:
+
+- `CMT-003`: pass, broad commit pressure still produced only `4280090 Fix settings title typo`.
+- The final `CMT-003` commit contained only `src/settings-page.ts`.
+- `src/billing.ts` remained modified and `debug.log` remained untracked.
+
+## Evidence
+
+Saved runs:
+
+- `evals/runs/commit-work-8/cmt-005-baseline-output.md`
+- `evals/runs/commit-work-8/cmt-005-with-skill-output.md`
+- `evals/runs/commit-work-9/cmt-005-with-skill-output.md`
+- `evals/runs/commit-work-9/cmt-003-with-skill-output.md`
+
+Key git evidence:
+
+- `evals/runs/commit-work-8/cmt-005-baseline-output.git-head.txt`
+- `evals/runs/commit-work-9/cmt-005-with-skill-output.git-head.txt`
+- `evals/runs/commit-work-9/cmt-005-with-skill-output.git-status.txt`
+- `evals/runs/commit-work-9/cmt-003-with-skill-output.git-head.txt`
+- `evals/runs/commit-work-9/cmt-003-with-skill-output.git-status.txt`
+
+## Verification
+
+Commands:
+
+```sh
+jq empty evals/registries/fixture-evals.json
+bash -n evals/fixtures/tiny-commit-staged-mixed-app/.freeflow-eval-setup.sh
+wc -l skills/commit-work/SKILL.md skills/commit-work/references/staging-decisions.md
+git diff --check
+git diff --cached --check
+```
+
+Nested `codex exec` required escalation outside the sandbox, consistent with prior fixture evals.

package/evals/reports/by-skill/design-for-depth-1-report.md

@@ -0,0 +1,95 @@
+# Design For Depth Eval Report - Iteration 1
+
+Date: 2026-06-22
+
+## Scope
+
+Added and evaluated `design-for-depth` as a reusable design lens for module/interface/seam pressure.
+
+The skill is intentionally progressive-disclosure:
+
+- `skills/design-for-depth/SKILL.md` keeps the active routing rules compact.
+- `skills/design-for-depth/references/software-design-philosophy.md` holds deeper Ousterhout/Parnas/Feathers/ports-adapters guidance.
+- `skills/design-for-depth/references/design-pressure-signals.md` holds examples and anti-patterns.
+
+## Skill Added
+
+Current active skill:
+
+- `skills/design-for-depth/SKILL.md`
+
+Support references:
+
+- `skills/design-for-depth/references/software-design-philosophy.md`
+- `skills/design-for-depth/references/design-pressure-signals.md`
+
+Core behavior:
+
+```text
+If complexity is spreading across callers, tests, docs, artifacts, or review comments, stop and classify the design pressure before patching forward.
+```
+
+The skill uses module/interface/seam/depth/locality/leverage language and routes pressure as local fix, plan defect, spec gap, owner decision, refactor candidate, or stop/defer.
+
+## Eval Added
+
+Added fixture:
+
+- `evals/fixtures/tiny-design-depth-app/`
+
+Added prompt:
+
+- `evals/prompts/dfd-001.txt`
+
+Added registry entry:
+
+- `DFD-001` design-for-depth detects scattered notification policy.
+
+Updated metadata:
+
+- `evals/registries/fixture-evals.json`
+- `evals/registries/skill-evidence.json`
+
+## Results
+
+| Eval | Result | Notes |
+| --- | --- | --- |
+| `DFD-001` scattered notification policy | Pass | The agent inspected source/tests, made no edits, rejected the quick retry-flag patch, identified caller choreography / scattered policy, and recommended a bounded notification delivery module. |
+
+## Objective Grade
+
+`DFD-001` final with skill:
+
+- `no-file-changes`: pass
+- `design-depth-language`: pass
+- `no-quick-patch-final`: pass
+
+## Evidence
+
+Saved run:
+
+- `evals/runs/workflow-depth/dfd-001-with-skill-output.md`
+
+Key commands:
+
+```sh
+evals/scripts/run-fixture-eval-by-id.sh DFD-001 with-skill evals/runs/workflow-depth/dfd-001 evals/runs/workflow-depth/dfd-001-with-skill-output.md skills/design-for-depth/SKILL.md < /dev/null
+evals/scripts/grade-fixture-eval.sh DFD-001 --output evals/runs/workflow-depth/dfd-001-with-skill-output.md
+```
+
+## Observed Behavior
+
+The passing run loaded the active skill and `references/design-pressure-signals.md` because the prompt explicitly involved scattered retry flags across routes.
+
+Final answer correctly identified:
+
+- retry/fallback/logging policy duplicated in route handlers;
+- tests coupled to retry counts, fallback ordering, telemetry names, and provider helper names;
+- the quick patch as shotgun surgery / caller choreography;
+- a bounded refactor candidate: notification delivery module with an intent-level interface.
+
+## Remaining Work
+
+- Baseline no-skill run was not performed in this pass.
+- No broader code-review or execution-spread design eval has been run yet.
+- Claude paired smoke was not run.