PR 5: cd-attribution + subshell isolation + bash -c recursion

IMPLEMENTATION_PLAN.md

@@ -114,20 +114,26 @@ bulldoze priorities.
 - [x] Flag-value path classification table (`git -C` is path; `curl -d`
       is body data; `docker -v` is single literal IsPath=false per #8)
 
-### 8. cd-in-compound propagation (SPEC §9)
-
-- [ ] First-clause `cd` sets attributed cwd for the compound
-- [ ] Subsequent clauses receive synthetic `Arg` with
-      `IsCwdAttribution=true`
-- [ ] Subsequent `cd` in the same compound replaces attribution
-- [ ] Subshell boundaries reset attribution
-
-### 9. Subshell + bash -c surfacing (SPEC §10)
-
-- [ ] Flatten subshell clauses into parent's `Clauses` with
-      `IsSubshell=true`
-- [ ] Surface `bash -c` inner clauses inline with `IsBashCWrapped=true`
-- [ ] Recursion-depth cap
+### 8. cd-in-compound propagation (SPEC §9) — PR 5, complete
+
+- [x] First-clause `cd`/`chdir` sets attributed cwd; only those two verbs
+      propagate (interp #5)
+- [x] Subsequent clauses receive synthetic `Arg` with
+      `IsCwdAttribution=true`; `Kind=Literal, IsPath=true` when cd
+      target resolved, `Kind=DynamicSkip, IsPath=false` when dynamic
+      (interp #6)
+- [x] Subsequent `cd` in the same compound replaces attribution
+- [x] Subshell boundaries isolate attribution via push/pop stack
+
+### 9. Subshell + bash -c surfacing (SPEC §10) — PR 5, complete
+
+- [x] Flatten subshell clauses into parent's `Clauses` with
+      `IsSubshell=true`; sibling subshells handled via SubshellStack IDs
+- [x] Surface `bash -c "..."` / `sh -c "..."` inner clauses inline with
+      `IsBashCWrapped=true`; outer cd attribution doesn't leak into
+      inner shell (v0.1 decision)
+- [x] Recursion depth cap at 5 → outer
+      `ParsedCommand.IsUnparseable=true` (interp #4)
 
 ### 10. Hand-authored corpus (SPEC §13 — minimum 105 entries)
 

SPEC.md

@@ -680,8 +680,15 @@ The agent's natural idiom is `cd /target && cmd1 && cmd2`. Bash semantics:
 
 ### Rules
 
-1. **First clause is a `cd`-family verb** (per §6.2): the cd target becomes
-   the **attributed cwd** for subsequent clauses in the same compound.
+1. **First clause is a `cd` or `chdir` verb**: the cd target becomes the
+   **attributed cwd** for subsequent clauses in the same compound. **Only
+   `cd` and `chdir`** propagate attribution per locked interpretation #5.
+   `pushd`, `popd`, `push-location`, and `set-location` are still listed
+   in `CwdVerbs` so their first non-flag positional is path-classified
+   (the target shows up as `IsPath=true`), but they do **not** add a
+   synthetic attribution arg to subsequent clauses. A future v0.1.x or
+   v0.2 with PowerShell support may model `pushd`/`popd` as a proper
+   directory stack.
 2. **Subsequent clauses inherit the attributed cwd** as if it were
    prepended with `-C` semantics. Specifically: a synthetic `Arg` with
    `IsPath=true`, `Resolved=<cd target>`, and `Kind=Literal` is added to
@@ -693,17 +700,40 @@ The agent's natural idiom is `cd /target && cmd1 && cmd2`. Bash semantics:
 
 3. **A subsequent `cd`** in the same compound **replaces** the attributed
    cwd for clauses after it. (`cd /a && cmd1 && cd /b && cmd2` → cmd1
-   inherits `/a`, cmd2 inherits `/b`.)
+   inherits `/a`, cmd2 inherits `/b`.) The replacing `cd /b` itself still
+   *receives* `/a` as a synthetic attribution arg (rule 2) before becoming
+   the new source — additive semantics per rule 5.
 
 4. **Subshell boundaries reset attribution.** `cd /a && (cd /b && cmd1) && cmd2`:
    cmd1 (inside subshell) inherits `/b`; cmd2 (outside subshell) inherits
-   `/a` (the subshell's `cd /b` does not leak out).
+   `/a` (the subshell's `cd /b` does not leak out). A subshell *inherits*
+   outer attribution on entry (so `cd /a && (cmd)` still attributes cmd
+   to /a) but its own cd changes stay isolated.
 
 5. **Attribution does not change the clause's verb or original args.**
    The attribution is purely additive — the `cd` clause itself is still
    parsed normally, and subsequent clauses retain everything the user
    typed, plus the synthetic Arg.
 
+### Dynamic-cd attribution (locked interpretation #6)
+
+When the cd target itself is `Kind=DynamicSkip` (e.g. `cd $REPO`), we
+statically don't know the resolved cwd. To preserve the cwd-uncertainty
+signal for subsequent clauses:
+
+- A synthetic `Arg { Raw="<dynamic-cwd>", Resolved=null, Kind=DynamicSkip,
+  IsPath=false, IsCwdAttribution=true }` is appended to each subsequent
+  clause (instead of the literal-cd flavor).
+- Relative path args in subsequent clauses are not re-resolved against a
+  fall-back cwd; they surface as `Kind=DynamicSkip, IsPath=false,
+  Resolved=null` so consumers route to safe-fail rather than trust a
+  guessed working directory.
+
+Consumers that iterate `IsPath=true` args won't see the synthetic
+attribution arg; consumers that specifically check `IsCwdAttribution`
+can detect "this clause's cwd context is unknown" and elevate to
+user-prompt instead of treating it like a default-cwd command.
+
 ### Example
 
 Input: `cd /target && git -C /other log && cat file.txt`
@@ -739,15 +769,11 @@ already see the resolved path in another arg.
 ### Subshells
 
 Subshells are clauses wrapped in parens: `(cd /a && cmd)`. The parser
-recognizes the parens, parses the inner command, and emits a single
-clause with:
-
-- `Verb` = empty (a subshell has no verb of its own)
-- `Args` = empty
-- `IsSubshell = true`
-- A nested `ParsedCommand` field — **but** since the AST should be flat
-  for consumer convenience, instead we **flatten** the subshell into the
-  parent's `Clauses` list with each inner clause's `IsSubshell=true`.
+recognizes the parens and **flattens** the subshell's inner clauses into
+the parent's `Clauses` list, marking each with `IsSubshell=true` so
+consumers can distinguish them from outer-compound clauses. A subshell
+*inherits* the outer compound's cd attribution on entry but its own cd
+changes stay isolated to the subshell (rule 4 above).
 
 Specifically: `(cd /b && cmd) && cmd2` produces three clauses:
 
@@ -779,7 +805,10 @@ by the recursion. Consumers that care that this came from a wrapper can
 inspect `IsBashCWrapped` on the surfaced clauses.
 
 **Recursion limit:** parse `bash -c "bash -c ..."` chains up to depth 5.
-Deeper nesting → mark the deepest clause as `IsUnparseable = true`.
+Deeper nesting → set the outer `ParsedCommand.IsUnparseable = true` with
+reason `"bash -c recursion depth exceeded (>5)"` per locked interpretation
+#4. (`Clause` has no `IsUnparseable` field; we surface the overflow on the
+top-level ParsedCommand so consumers safe-fail per §11.)
 
 ---
 

openspec/changes/v0.1-locked-interpretations/tasks.md

@@ -161,23 +161,61 @@ the SPEC.md sections that get updated alongside the implementation.
 
 ## 5. PR 5 — cd attribution + subshells + bash -c (interpretations #4, #5, #6)
 
-- [ ] 5.1 `Internal/Bash/Parsing/CdAttributionContext.cs` (parser-internal
-      mutable; output AST stays immutable)
-- [ ] 5.2 Only `cd`/`chdir` propagate (interpretation #5); pushd/popd
-      parse but don't propagate
-- [ ] 5.3 Synthetic `Arg{ IsCwdAttribution=true, … }` appended to
-      subsequent clauses; `Kind=DynamicSkip` when cd target is dynamic
-      (interpretation #6)
-- [ ] 5.4 Subshell `(...)` parsing with attribution stack push/pop
-- [ ] 5.5 `bash -c "..."` / `sh -c "..."` recursion (cap at 5)
-- [ ] 5.6 Recursion overflow → outer `ParsedCommand.IsUnparseable=true`
-      (interpretation #4)
-- [ ] 5.7 Update `SPEC.md` §9 (clarify which CwdVerbs propagate; add
-      dynamic-cd attribution subsection); §10 (replace "mark the deepest
-      clause" wording with "set ParsedCommand.IsUnparseable=true")
-- [ ] 5.8 File GitHub issue: "Model pushd/popd directory stack semantics"
-- [ ] 5.9 Open OpenSpec change `cd-attribution-clarifications` for the
-      §9/§10 deltas
+- [x] 5.1 `Internal/Bash/Parsing/CdAttributionContext.cs` — parser-internal
+      mutable; SubshellStack of monotonic IDs (handles sibling subshells
+      cleanly); SetLiteralAttribution / SetDynamicAttribution; HasAttribution
+- [x] 5.2 Only `cd`/`chdir` propagate (interpretation #5); pushd/popd
+      parse as CwdVerbs but don't propagate. Test
+      `Pushd_does_not_propagate_attribution` pins this.
+- [x] 5.3 Synthetic `Arg{ IsCwdAttribution=true, … }` appended to
+      subsequent clauses. `Kind=Literal, IsPath=true, Resolved=<cwd>`
+      when cd target resolved; `Kind=DynamicSkip, IsPath=false,
+      Resolved=null, Raw="<dynamic-cwd>"` when cd target was DynamicSkip
+      per interpretation #6.
+- [x] 5.4 Subshell `(...)` parsing with attribution stack push/pop.
+      `cd /a && (cd /b && cmd1) && cmd2` — cmd1 sees /b attribution;
+      cmd2 sees /a (subshell's /b doesn't leak out). `IsSubshell=true`
+      set on every clause inside a subshell.
+- [x] 5.5 `bash -c "..."` / `sh -c "..."` recursion replaces PR 3's
+      single-clause framework. Outer `bash -c` consumed; inner clauses
+      surfaced inline with `IsBashCWrapped=true`. Outer cd attribution
+      does NOT leak into inner shell (v0.1 decision).
+- [x] 5.6 Recursion depth cap at 5 → outer
+      `ParsedCommand.IsUnparseable=true` with reason `"bash -c recursion
+      depth exceeded (>5)"` per interpretation #4.
+- [x] 5.7 SPEC §9 updated (rule 3 explicit on which verbs propagate;
+      dynamic-cd attribution subsection added per interp #6); §10
+      updated (subshell flattening rephrased; bash -c recursion-limit
+      replaced with "set ParsedCommand.IsUnparseable" per interp #4).
+- [x] 5.8 BashResolver internal overload `Resolve(raw, treatAsPath,
+      options, workingDirectoryUnknown)` lets the propagator force
+      DynamicSkip on relative-path args under dynamic-cd (interp #6)
+      without polluting the public `BashParserOptions` surface.
+- [x] 5.9 30 new corpus entries (71-100): 10 cd-in-compound + 10
+      subshell + 10 bash -c. 5 PR 3-4 corpus entries refreshed
+      (25, 28, 34, 35, 52).
+- [x] 5.10 8 new parser tests covering attribution propagation, subshell
+       isolation, sequential cd, pushd non-propagation, dynamic cd,
+       sibling subshells, bash -c depth-2 success, depth-6 overflow.
+- [x] 5.11 **337/337 tests passing** (was 296 at PR 4 baseline). Public
+       API surface unchanged.
+- [ ] 5.12 File GitHub issue: "Model pushd/popd directory stack semantics"
+       (interp #5 option-C upgrade)
+
+### PR 5 follow-ups (tracked for PR 6)
+
+- Possible SPEC clarification: pin the sentinel `Raw` value for
+  dynamic-cd synthetic attribution args (current implementation uses
+  `"<dynamic-cwd>"` but SPEC is silent).
+- Outer cd attribution into bash -c inner clauses (v0.1: doesn't
+  propagate; v0.1.x or v0.2 may revisit).
+- `cd -` (jump-to-previous-dir): currently treated as "no attribution
+  change" because `-` is detected as a flag (no non-flag positional).
+  v0.1.x may want explicit handling.
+- Existing PR 3 corpus entries 21-24, 26-27, 29-33 (the compound
+  entries that had no `cd` prefix) likely don't need updates; but PR 6
+  should audit all 100 corpus entries for AST drift now that PRs 1-5
+  have all landed.
 
 ## 6. PR 6 — Corpus completeness + PII audit (interpretation #7)