TestBuildToolsExposesExactlyFourTools pins the tool roster: bash, read_file, write_file, edit_file, in that order, with no loop/control tool. Order is

TestCRLFLineEndings: Windows \r\n counts each line once, not twice.

TestCancelMidStreamPreservesStreamedText: Ctrl+C while content is streaming keeps the partial response visible (flushed to scroll) and appends the can

TestCarriageReturnLineEndings: lone-\r separators (old-mac-style, some VS Code TERM setups) must still yield a correct chip line count. bubbles/textar

TestChat401: maps to typed ErrUnauthorized.

TestChat401DrainsBodyForConnReuse: a 401 carrying a body must have that body drained before close, or Go's transport discards the TCP connection inste

TestChat402: budget exhaustion surfaces as a typed error with the snapshot reporting zero remaining, so the UI paints the depleted state immediately.

TestChatDispatchesToolCallsOnFinishStop: Ollama's /v1 shim sometimes emits finish_reason="stop" even after streaming tool_calls. The client must still

TestChatDoesNotFallBackOnUnrelatedThinking: a 400 that is NOT about reasoning but happens to contain both "not support" and the word "thinking" must n

TestChatFallsBackWhenOllamaRejectsThinking: Ollama rejects reasoning_effort on non-thinking models with a 400 saying `<model> does not support thinkin

TestChatFallsBackWhenReasoningEffortRejected: newer OpenAI models reject tools + reasoning_effort on /v1/chat/completions with a 400. postChat must dr

TestChatIdleTimeoutAbortsStalledStream reproduces the exact hang: the server returns 200 OK then sends nothing. Without the idle watchdog scanner.Scan

TestChatIdleWatchdogResetByFrames pins that an alive-but-slow stream is NOT aborted: frames spaced under the idle window each reset the watchdog, so a

TestChatMidStreamErrorFrameSurfacesAsError: OpenAI-compatible backends (and OpenRouter-style proxies like the hosted endpoint) report a post-200 provi

TestChatOtherHTTPError: non-2xx (not 401/402) surfaces as a generic error carrying only the first body line.

TestChatReadsUsageTokens: tokens come from `usage.completion_tokens` (and prompt_tokens rides along for the debug-log calibration), not content length

TestChatSendsStreamIncludeUsage: servers emit the usage block only when `stream_options.include_usage:true` is in the request; without it the per-turn

TestChatStreamsContent: content deltas merge into one final string.

TestChatStructuredErrorFallsBackToMessage: with only `error.message` (no provider_hint), surface that, not the raw JSON envelope.

TestChatStructuredErrorPrefersProviderHint: the hamrpass proxy wraps upstream errors as `{"error":{"message":...,"provider_hint":...}}`. The client mu

TestChatToolArgsStreamLive: each tool-call arguments fragment is forwarded as an EventToolArgs as it arrives, so the UI can tick its live token estima

TestChatToolCall: tool_calls in a delta emit EventToolCall and ride along in EventDone.Final.ToolCalls so the next turn can replay the assistant messa

TestChatToolCallFragmentedArgs: `arguments` arrives as JSON fragments, each invalid alone. The client must accumulate raw and parse once at finish_rea

TestChatToolCallLateIDPreserved: spec ships the tool_call `id` in the first fragment, but a sloppy provider may delay it. The client must update slot.

TestChatToolCallMalformedArgsPreservesMarker: on invalid `arguments` JSON (provider bug), the client surfaces a sentinel key rather than an empty args

TestChatToolCallMultipleByIndex: two tool calls interleaved across chunks. Each fragment must route to its slot by `index`, not by slice position.

TestChatUnreachable: transport failure surfaces as ErrUnreachable.

TestCheckHonoursEnvDisableFlag: CODEHAMR_NO_UPDATE_CHECK=1 must short-circuit Check before any HTTP work, sparing CI/offline launches the fetch deadli

TestCheckReportsStale: published hash differs from local → Check returns true (update available), driving the maybeSelfUpdate trigger.

TestCheckReportsUpToDate: local hash matches the manifest → Check returns false. The steady-state path that keeps the spinner quiet after a release.

TestCleanupOldNoopWhenMissing: cleanup must be silent (no error/log/panic) when there is no .old file, the steady state once an update has settled.

TestCleanupOldRemovesStaleFile: the previous Apply's .old must be removed at the next launch so it doesn't accumulate. On Windows .old stays locked un

TestCleanupOldSweepsOrphanedTempFiles: a Ctrl+C mid-download kills the process before Apply's deferred Remove runs (no signal handler exists that earl

TestClearResetsFailureStreak: /clear starts the conversation over, including the repeated-failure backstop's counters.

TestClearResetsSessionTokens: /clear wipes the conversation AND the session counter: starting over from zero is the whole point.

TestClearWipesQueue: /clear resets the conversation, so a queued follow-up must go with it (it would target a conversation that no longer exists).

TestConfigFilePermissionsAreOwnerOnly is the regression for a world-readable hamrpass key: /hamrpass stores the bearer token in plaintext, so any loca

TestCtrlCCancelsInflightOp: with a turn in flight, Ctrl+C cancels, clears pending tool calls, and leaves a "✗ cancelled" line in scrollback.

TestCtrlCIdleArmsThenQuits: first Ctrl+C in idle arms the status bar and returns a Tick cmd; second Ctrl+C before the window expires returns Quit.

TestCtrlCPopoverClosesInsteadOfQuitting: with the popover open and no in-flight op, Ctrl+C dismisses the popover and does not arm quit.

TestCtrlDEmptyQuits: Ctrl+D on empty textarea returns a Quit command.

TestCtrlDMidTurnDoesNotQuit: the textarea is empty during a running turn (submit resets it), so without the phase gate a reflexive Ctrl+D would quit i

TestCtrlDNonEmptyNoOp: Ctrl+D with text in the textarea is a no-op: no quit and no character deletion.

TestCtrlLClearsPromptNotScrollback: Ctrl+L matches Claude Code: it clears the typed input and forces a terminal redraw, but conversation scrollback st

TestCtrlLClosesPopover: Ctrl+L clears the popover along with the prompt. Left open on stale suggestions, the next Enter would take the has-selection p

TestCtxPressureTripwire: the round_done companion warning fires only when the server-counted prompt reaches 95% of the active window. The char/4 packe

TestDamagedLabelAmongIdenticalChipsDropsWholeGroup: two pastes with the same line count render identical labels. A word-delete at the boundary (Ctrl+W

TestDamagedLabelUniqueChipDropsOnlyItself: the group-drop above must not over-trigger; a damaged label with no twin drops alone and the other chip kee

TestDebugLogFilePermsAreOwnerOnly: the log captures every prompt and tool-call payload: bash args can carry heredoc secrets, so a world-readable log l

TestDeleteAtChipStartRemovesWholeChip: forward-Delete with the cursor right before the label removes the whole chip.

TestEditFileOverlappingOldString: strings.Count sees only one non-overlapping "==" in "a === b", but it matches at two positions with different result

TestEditFileRefusesNonRegular: edit_file reads the target first, so it needs the same guard as read_file.

TestEditFileWhitespaceNearMissHint: a miss whose only difference is indentation gets the diagnostic hint, and the file is left untouched (the hint is

TestEmbeddedPromptFitsFixedSystem guards the invariant the packer relies on: FixedSystem must reserve enough budget for the embedded system prompt PLU

TestEmptyReplyNudgeFiresOnceThenSurfaces: if the model stays empty even after the re-prompt (e.g. a server that deterministically swallows the call),

TestEmptyReplyNudgeReArmsAfterProgress pins the galaxy1 fix: once an empty reply has nudged this turn, a round that issues a real tool call is genuine

TestEmptyReplyNudgeRePromptsThenRecovers: a turn whose first round comes back with no content and no tool call (the dominant silent-death: a thinking

TestEmptyRunesKeyIsDropped: KeyMsg{Type: KeyRunes, Runes: nil} arises when bubbletea's parser chokes on a partial escape sequence. Falling through ins

TestEndTurnResetsPendingSoStaleCallsDoNotLeakIntoNextTurn: a turn can end with calls still in m.pending. Without resetting pending in endTurn, the nex

TestEndTurnResetsToolRounds: the runaway counter is per-turn, so endTurn must zero it or the next turn inherits a head start toward the cap.

TestEndTurnResetsVerifyNudged: the latch is per-turn, so endTurn must clear it or a later turn never re-grounds.

TestEnsureHamrpassLazyCreates: with hamrpass hidden from config.yaml, EnsureHamrpass returns a profile from the canonical seed values, and is idempote

TestEntryRestoreRoundTrip: Entry()+Restore() fully recovers chip state: display text, spans, expanded Value(). Backs ↑/↓ history replay.

TestErrorMessageBudgetExhausted: 402 produces the depleted hint pointing users at the top-up page rather than a stack-trace style wrap.

TestErrorMessageUnauthorized: 401 names the rejected key and the exact config path so the user can fix it without guessing.

TestErrorMessageUnreachable verifies the unreachable hint names the active profile's URL and steers the user toward /models to switch profiles.

TestEscFromCommandLevelClosesAndClears: Esc at command-level closes the popover AND clears the textarea, so the user returns to a blank prompt. Typing

TestEventDoneFlushesStreamingThroughGlamour: EventDone moves raw streamed text into the committed scroll via the Markdown renderer and empties the str

TestEventErrorPreservesStreamedText: a stream error mid-content flushes the partial text before appending the error line, same principle as cancel, us

TestExecuteEditFileRefusesMissingNewString: same guard as write_file's content - a dropped new_string must not decode to "" and silently delete the ma

TestExecuteWriteFileRefusesMissingContent: valid-JSON args with no string "content" ({"path": ...} alone, or "content": null) must not decode to "" an

TestFetchHashHandlesCorruptManifest: a manifest not in `<hash> <name>` form must not crash fetchHash; it just yields an empty hash.

TestFinalizeFoldsInFlightEstimate: a turn aborted mid-stream (Ctrl+C, error) has no EventDone to fold the current round's tokens into turnTokens; they

TestFirstResizeDoesNotClearScreen: the first WindowSizeMsg must not ClearScreen; the terminal still holds the user's pre-launch shell output; wiping i

TestFocusBlurMsgsAreInert: terminal focus reports arrive as tea.FocusMsg / tea.BlurMsg under tea.WithReportFocus. They must never touch the textarea (

TestFromHeadersRejectsNaNAndInf: ParseFloat accepts NaN/±Inf without error and the [0,1] clamp is a no-op against NaN, so they'd render as int(NaN*100

TestHamrpassLazyCreatesProfile: a user who hid hamrpass from config.yaml can still activate it by pasting a key. /hamrpass <key> creates the profile f

TestHamrpassNoArgsShowsExplainerWhenUnset: `/hamrpass` with no key set prints the guided block, including the unset status line and the purchase URL.

TestHamrpassNoArgsShowsSetWhenKeyPresent: status line flips to `set` when the hamrpass profile already has a key.

TestHamrpassRejectsControlChars: a key with an embedded escape / NUL / DEL must be rejected by hamrpassValidate, not persisted, else every dial-out er

TestHamrpassRejectsMultipleArgs: a paste with embedded whitespace splits into multiple args via strings.Fields. The handler refuses with a dedicated m